On Topline, Tim Rutten said real bank AI deployments are almost 99% deterministic, with the language model reasoning at only one heavily guardrailed moment.
He says that once AI has been trained with skills, business context and information, a price should come from a deterministic method of inputs, formulas and math. He argues computation is cheap, so the system can produce several perspectives that a team can evaluate. Whether a team accepts the output is still up to its own judgment.
“you don't let it when it comes to pricing right you don't let it guess you don't let it go and search the web and give me the answer”
Keith separates the deterministic numbers, which are handled in code, from the qualitative insight LLMs add. His example is looking past last quarter's closed count to ask how many deals nearly closed or slipped because of missing features or competitors.
“how many of them were sort of on the cusp of actually being able to close because of missing features, your competition and so forth.”
Asad Zaman challenged whether humans checking every AI-generated forecast formula saves any time. Keith said the work was a one-time implementation: Lightfield's head of finance reviewed every formula and some prompts were converted to code. After that the team makes only fine adjustments week over week. Keith acknowledged 'some might disagree' but said LLMs are not the right tool for math.
“We converted some things from prompts to real code. I don't want LLMs doing math. I want real code doing math.”
He said banks do not apply AI all the way through a workflow. Instead, they use a language model for a single interpretation at the right point, heavily guardrailed, while the rest of the process is rule-based. He presented this as the way banks get to an acceptable error rate.
“the actual implementations are almost 99% deterministic, and there's this one little percent that actually gets them to gain, to reason at the right moment in time.”
He says there is a misunderstanding that everything must go through an LLM. In GTMOS most of the system is deterministic code, with automation and workflows around certain activities, and LLM calls reached through an AI gateway only where needed. He frames the question as finding the right balance.
“I would say that almost almost 70 to 80% of GTMos is not LLM based. It's actually very deterministic”
Kyle's test is to define the business challenge, list the ways to solve it, and use AI when the work needs a large amount of data ingested and a non-deterministic judgment, such as pattern recognition. He said an agent at every step often leads to sloppy outputs.
“I think people too often it's a hammer looking for a nail and it's like what can I do with AI? And that's fundamentally sort of the wrong approach.”
He gave the example of deciding whether a prospect's tone is right to ask for a credit card, which he said AI should not do. He said the tools lack the capability for tone analysis. He said agents do well when tasks are split into deterministic choices and when they handle non-deterministic inputs such as a website.
“tone analysis is like something that, you know, the tools don't have the capability to do”
Wade said these workflows look like traditional Zapier workflows with an LLM step inserted to handle unstructured data, make a decision, summarize, or categorize. LLMs do well with lots of context, so deterministic steps can feed precise context and target the model at a specific task. He said this gives higher reliability for the whole flow.
“But then in the middle of it, you're inserting either an LLM step or, you know, an agentic step that handles that handles some like unstructured data”
On Topline, Kyle Norton said giving every Owner.com employee AI accounts did not produce results, while systems built centrally by a small group and deployed into Salesforce did.
He said the centralized model should create playgrounds for different functions, with constraints so that any question about revenue returns consistent data based on the user's role and security access. He also said that teams given free rein burned tokens quickly and were inefficient, while centralizing lets the company optimize spend against output.
“I think ultimately the answer is a centralized model that creates these playgrounds for different functions”
He described a dinner conversation with other operators where people had given every team access to agent tools, and in a meeting each group's data set, built with its own agents, did not match. He said the meeting became a discussion of where the data came from rather than what the company should do.
“none of the data matches”
Rutten says GTMOS runs on enterprise authentication, so people log in with their own entitlements, and the Salesforce, research and email connections are made once at enterprise level, metered so no one can pull the full CRM at once. He says the biggest hurdle in building it was getting compliance sign-off for a non-regulated business that sells to regulated banks.
“we basically do that at a enterprise level so we have these tools under the hood fully connected fully guarded and metered”
Alex Bilmes said that when each of a thousand reps has a different value proposition, sales process and way of describing value to an ICP, coordination and momentum suffer. He argued for combining customer data, methodology, enablement assets and messaging into one foundation that any human or agent queries, so answers align with corporate strategy.
“Centralization is a lot more important.”
Alex Bilmes described a customer where a rep communicated a whole new set of SKUs and a new product offering, with pricing and packaging, that the company did not have. He used it to argue that individual AI answers need to be replaced by a centralized, company-approved source.
“We had a customer where a rep communicated a whole new set of SKUs. and a totally new product offering to the customer, because ChatGPT told them to.”
Greg Casale says giving each rep an AI sourcing tool does not produce the value; instead, the team pulls ICP, TAM accounts, contacts, phone numbers and propensity signals into the back office. He says the prepared data is sent to SDRs through Sales Loft, which removes human sourcing errors such as the right person at the wrong company. He says that research which used to take hours a day is now already in place when a rep makes the call.
“instead of arming each rep with an AI based sourcing tool so they can get really smart about who they're calling that's not that's not the value”
Kyle Norton says his company's centralized approach emerged early, when a small group became obsessed with applying AI. They built things and deployed them to reps inside the surfaces they already used rather than asking everyone to learn new tools. He says the central owner need not be one person or an AI tsar and could sit in engineering with applied AI or in RevOps.
“we just built things, deployed it to the reps in the surfaces that they already use”
A host says decentralized AI gives the consumer benefits that individuals feel, such as emails written faster, meeting notes and decks built faster. He says it does not produce use cases that can be shown to move the business in a substantial way and are ready for production. He describes the output quality as depending on each person's taste.
“But you don't get like real business use cases, like the ones that you can say, ah, this move the needle for the business in a really substantial manner”
Kyle Norton says the popular advice is to give everyone Claude accounts and have them build, and he pushed AI adoption heavily about 12 months ago. In practice, though, the implementations that worked were highly centralized. A small group (the VP of BizOps and data, the VP of RevOps and Kyle) drove the work, and the VP of BizOps and data did most of the building. The tools were then deployed into the tools reps already used, such as Salesforce and Salesloft, rather than as new apps.
“what we really found in practice is that all implementations were all highly centralized.”
On Topline, Sam Jacobs said AI's primary feature is persuasiveness and certainty, not accuracy, which is why subject-matter experts must say when an output is wrong.
Danfoss has built a system that helps sales teams work smarter, has more advanced analytics, and is building its own pricing software with AI embedded to help users set the right prices. Jacob says AI is very good at using existing data and extrapolating from it, so humans must control it and add commercial judgment it cannot generate itself.
“we have quite good master data which is really fundamental to make AI work.”
He describes the scar tissue from implementing approaches that did not work, which lets experienced pricers recognise when an AI suggestion is unlikely to work. He says training AI takes time that not everyone has, and encourages the next generation to still learn what good looks like by making the mistakes experienced pricers made.
“And we know what good looks like, right?”
Mark Stiving says AI is wrong on pricing questions he knows the answers to, yet he believes it on unfamiliar topics such as a skin rash, which he calls dangerous. He says he tries to use AI as a first-pass answer and does not necessarily trust it.
“I know that it's wrong because I happen to know the answers.”
Responding to Mike, Sam said the primary feature of AI is persuasiveness and certainty, not necessarily accuracy. He said that is why subject-matter experts are needed to say when an output does not work or is not correct.
“The primary feature of AI is persuasiveness and is certainty. It is not necessarily accuracy.”
She compares AI-built models to an FP&A analyst's Excel model with one mistake, which can flow through a whole bookings model. She says a human still needs to check the work, and that AI makes human judgement on trade-offs more important rather than less.
“I worry about it if you're relying on it solely to give you the answer without the human review.”
Diane Wu says AI is a strong source of raw information and analysis but needs a CSM or CS leader with curiosity and technical fluency to prompt it well and ask the right discovery questions. She says this human judgment is what turns output into differentiated value. She frames the human role as curating and adding context, not replacing the CSM.
“when you use an AI tool, never ever take it for face value, right?”
Nick said he has seen teams damaged by using generative AI on top of complex models after they made mistakes by trusting what the bot said. Because Dreamdata's customers take its numbers to their boards and C-suites, he said the company approaches AI carefully so its output is accurate enough for decisions.
“I've seen teams wiped out utilizing generative AI on top of these complex models completely because they made mistakes by just trusting what the the the the bot saying.”
On [Un]Churned, Margo Martin said Deltek's in-house AI support tool worked, but buying the same thing avoided needing a team to maintain it as models changed.
Potter says organizations shifting to homegrown AI agents and Slack-based interfaces must also build how they maintain, version and iterate on those tools, and plan for a system crash at 3 a.m. He compares it to earlier cycles when companies built their own mail, documentation, procurement and data warehouses and later handed them to vendors, and says some companies are already pulling back.
“when you start to build your own engineering aspect, you also have to build the way that you're going to maintain those, the way that you're going to version those, the way that you're going to iterate upon those.”
Snehal said organisations that try to keep updating models, workflows and MCPs themselves may spend three to five times more resources. He said a proof of concept or MVP might work for two or three months but then become outdated. He said customers who tried building this way later came back looking for a solution to the problem.
“you're spending your your spending 3x or 5x more resources, right? And you will be outdated in six months or so.”
About three years earlier, Deltek weighed building and buying AI capabilities for support and ran both in parallel. Margo Martin said the tool they built worked and could have been rolled out, but they could buy basically the same thing without a team to keep it running as the models changed. Since then she said Deltek has mostly bought, while still playing around with building to some extent.
“we could buy basically the same thing. And then we didn't have to worry about keeping it up to date, right? The models were changing.”
His list: an AI model building the data model and the pricing and rules engine from scratch with good opinions, and an agent architecture on top. Beyond that, the tool must serve IT, deal desk, RevOps, sales and finance workflows, integrate with billing and CLM so amendments and renewals are right, and handle the legal risk of sending wrong quotes. It must also be load-tested for thousands of salespeople, maintained as people leave, and reworked when the CEO wants a new token-based pricing model in two months. He adds that a model drawing on prior CPQ tools will reproduce 'some tool designed for IT and RevOps that sales absolutely hates.'
“CPQ is unique because it touches IT, deals desk, revops, sales, finance, minimally, just to get from a quote to a deal approved to the customer.”
He argued that a company building a highly specialized piece of software with general AI tools will not get something better, cheaper or easier to maintain than an off-the-shelf product, because the risk and benefit are not in line. He expects specialized AI applications to win in customer success and support, where problems are smaller, tightly controlled and have data to train models.
“a lot of the big winners that are going to be out there over the next few years are going to be AI application companies”
Ranked by how many independent voices make each point and how specific their evidence is. Co-hosts of a show count as one voice, and a point needs at least two shows to appear here.
said Justin Shriber (Topline), Tim Rutten (The Revenue Leadership Podcast), Alex Bilmes (Revenue Builders), Greg Casale (Revenue Builders), Kyle Norton (The Revenue Leadership Podcast, Topline)
He said the centralized model should create playgrounds for different functions, with constraints so that any question about revenue returns consistent data based on the user's role and security access. He also said that teams given free rein burned tokens quickly and were inefficient, while centralizing lets the company optimize spend against output.
“I think ultimately the answer is a centralized model that creates these playgrounds for different functions”
He described a dinner conversation with other operators where people had given every team access to agent tools, and in a meeting each group's data set, built with its own agents, did not match. He said the meeting became a discussion of where the data came from rather than what the company should do.
“none of the data matches”
Rutten says GTMOS runs on enterprise authentication, so people log in with their own entitlements, and the Salesforce, research and email connections are made once at enterprise level, metered so no one can pull the full CRM at once. He says the biggest hurdle in building it was getting compliance sign-off for a non-regulated business that sells to regulated banks.
“we basically do that at a enterprise level so we have these tools under the hood fully connected fully guarded and metered”
Alex Bilmes said that when each of a thousand reps has a different value proposition, sales process and way of describing value to an ICP, coordination and momentum suffer. He argued for combining customer data, methodology, enablement assets and messaging into one foundation that any human or agent queries, so answers align with corporate strategy.
“Centralization is a lot more important.”
Alex Bilmes described a customer where a rep communicated a whole new set of SKUs and a new product offering, with pricing and packaging, that the company did not have. He used it to argue that individual AI answers need to be replaced by a centralized, company-approved source.
“We had a customer where a rep communicated a whole new set of SKUs. and a totally new product offering to the customer, because ChatGPT told them to.”
Greg Casale says giving each rep an AI sourcing tool does not produce the value; instead, the team pulls ICP, TAM accounts, contacts, phone numbers and propensity signals into the back office. He says the prepared data is sent to SDRs through Sales Loft, which removes human sourcing errors such as the right person at the wrong company. He says that research which used to take hours a day is now already in place when a rep makes the call.
“instead of arming each rep with an AI based sourcing tool so they can get really smart about who they're calling that's not that's not the value”
Kyle Norton says his company's centralized approach emerged early, when a small group became obsessed with applying AI. They built things and deployed them to reps inside the surfaces they already used rather than asking everyone to learn new tools. He says the central owner need not be one person or an AI tsar and could sit in engineering with applied AI or in RevOps.
“we just built things, deployed it to the reps in the surfaces that they already use”
A host says decentralized AI gives the consumer benefits that individuals feel, such as emails written faster, meeting notes and decks built faster. He says it does not produce use cases that can be shown to move the business in a substantial way and are ready for production. He describes the output quality as depending on each person's taste.
“But you don't get like real business use cases, like the ones that you can say, ah, this move the needle for the business in a really substantial manner”
Kyle Norton says the popular advice is to give everyone Claude accounts and have them build, and he pushed AI adoption heavily about 12 months ago. In practice, though, the implementations that worked were highly centralized. A small group (the VP of BizOps and data, the VP of RevOps and Kyle) drove the work, and the VP of BizOps and data did most of the building. The tools were then deployed into the tools reps already used, such as Salesforce and Salesloft, rather than as new apps.
“what we really found in practice is that all implementations were all highly centralized.”
said Jonathan Moss (The Revenue Leadership Podcast), Jared Collins ([Un]Churned), Guy Galon ([Un]Churned), Josh Schachter ([Un]Churned)
Moss says the fix was, first, access for everyone and, second, tools connected to the right systems and context on AgentCore. He notes that the agent is the easiest part to build once the framework, tools and context exist. He also hedges that being a regulated healthcare company probably helped limit rogue building.
“We only had a set group or a set number of people that had it. And what we learned was, A, it pissed a bunch of people off”
Collins said he has encouraged his team to experiment, asking what repetitive day-to-day tasks could be turned into a one-click action, and how AI might surface insights from a customer contract. He said he is only doing some Excel vibe coding so far, and that individuals have already produced useful results in a few minutes.
“It doesn't have to be these massive kind of enterprise wide kind of initiatives to get that AI efficiency. It starts with just day-to-day work.”
He says the team should not set big targets but should solve a small use case first and see that it works. He encourages this because he wants Obrela's AI next steps to come from his team, not because he tells them to adopt them.
“So don't try to set big targets. Try to solve a small use case first and see that it works for us.”
He says every organisation has people, perhaps an ops person or a CSM, who are tinkering with AI on their own. He advises making them the poster child of the organisation, encouraging them and feeding them the resources they need. He says you have to promote AI adoption when it comes organically from those people. Kristi agreed.
“you can't always force it. You have to really kind of like promote it when it's coming organically from from those.”
Centralists focus on production systems that touch shared revenue data and security, while bottom-up advocates focus on building individual habits and buy-in.
said Rob Potter (Revenue Builders), Snehal Nimje (Topline), Margo Martin ([Un]Churned), Joubin Mirzadegan (Grit), Jim Richmond ([Un]Churned)
Potter says organizations shifting to homegrown AI agents and Slack-based interfaces must also build how they maintain, version and iterate on those tools, and plan for a system crash at 3 a.m. He compares it to earlier cycles when companies built their own mail, documentation, procurement and data warehouses and later handed them to vendors, and says some companies are already pulling back.
“when you start to build your own engineering aspect, you also have to build the way that you're going to maintain those, the way that you're going to version those, the way that you're going to iterate upon those.”
Snehal said organisations that try to keep updating models, workflows and MCPs themselves may spend three to five times more resources. He said a proof of concept or MVP might work for two or three months but then become outdated. He said customers who tried building this way later came back looking for a solution to the problem.
“you're spending your your spending 3x or 5x more resources, right? And you will be outdated in six months or so.”
About three years earlier, Deltek weighed building and buying AI capabilities for support and ran both in parallel. Margo Martin said the tool they built worked and could have been rolled out, but they could buy basically the same thing without a team to keep it running as the models changed. Since then she said Deltek has mostly bought, while still playing around with building to some extent.
“we could buy basically the same thing. And then we didn't have to worry about keeping it up to date, right? The models were changing.”
His list: an AI model building the data model and the pricing and rules engine from scratch with good opinions, and an agent architecture on top. Beyond that, the tool must serve IT, deal desk, RevOps, sales and finance workflows, integrate with billing and CLM so amendments and renewals are right, and handle the legal risk of sending wrong quotes. It must also be load-tested for thousands of salespeople, maintained as people leave, and reworked when the CEO wants a new token-based pricing model in two months. He adds that a model drawing on prior CPQ tools will reproduce 'some tool designed for IT and RevOps that sales absolutely hates.'
“CPQ is unique because it touches IT, deals desk, revops, sales, finance, minimally, just to get from a quote to a deal approved to the customer.”
He argued that a company building a highly specialized piece of software with general AI tools will not get something better, cheaper or easier to maintain than an off-the-shelf product, because the risk and benefit are not in line. He expects specialized AI applications to win in customer success and support, where problems are smaller, tightly controlled and have data to train models.
“a lot of the big winners that are going to be out there over the next few years are going to be AI application companies”
said Jonathan Moss (The Revenue Leadership Podcast), Tim Rutten (The Revenue Leadership Podcast), Kyle Norton (Topline)
His two objections were lock-in, which he sees as a big risk while things are moving fast, and giving a vendor the company's business logic, processes, context and IP. He says some parts of the stack should be bought, but the brain and harness are where to spend time building. Norton noted that the in-house technical capability to do this is not the norm in his experience.
“we were about to put all of this into a third party. And I was, and, and I was like, whoa, whoa, whoa, before we do that, how Are you going to port it over to something else now?”
Backbase did not renew 6sense because it did not fit its motion. Since 6sense informed campaigning and parts of the engine, they had to rebuild within two months. Rutten says the team built a white-box signal engine with a scoring rubric they tune themselves, which he says gives them go-to-market alpha once it connects to other building blocks.
“we completely built a signal engine bespoke that is not blackbox but white box I know exactly what is happening there”
Kyle Norton says the core intelligence of the organization needs to be built and managed centrally, and he doesn't think it should be bought. He says the things to buy are workflows and user experiences. He says build-versus-buy was a big topic of conversation at the Clay CRO summit he attended.
“I think you have to own the intelligence. You have to build the intelligence internally and buy things that are like workflows and user experience.”
Builders keep in-house only the proprietary company context and core intelligence, while buyers are weighing commodity or widely used tools whose upkeep outweighs customisation.
said Ian Tickle (The Revenue Leadership Podcast), Parag Agarwal (Grit), Chael Banks ([Un]Churned), Brad Casemore ([Un]Churned), Jared Collins ([Un]Churned)
Ian says being quicker is not the same as being more efficient or effective, and gains that aren't redeployed wash out. His example is automating reporting so that the analysts who build reports can work on programs or ICP and TAM analysis instead. He frames the goal as using tools to move people onto higher-value work, not just to make jobs easier.
“what do you do in that efficiency to make value out of it? Otherwise there's like a net-net, it washes out at the back end. We're just being quicker but we're not necessarily more efficient or effective”
In what he calls a semi-hypothetical example, a PE firm that once used gut feel to narrow buyout targets to four or five could use Parallel and models to run near-exhaustive 'simulations' across criteria before choosing. Because capital and people limit how many deals the firm can do, picking better creates enough value to justify heavy token spend. He calls this the next year of growth: 'work that was not happening yesterday.'
“It's going back to what does it take to make it 1000 x, right? It is work that was not happening yesterday.”
His examples: backend engineers doing more frontend work, frontend engineers changing APIs, and people using agents to run security reviews or push optimisations they would otherwise never have done. He sees true value in people taking more end-to-end ownership and pushing their own learning curves.
“it's actually people expanding their span of influence and ownership on the product. That's where I think there is true business value being generated in my mind.”
Chael says Okta uses AI to remove low-value tasks so people can have higher-level conversations, and that the efficiency gains are in service of effectiveness. He says the time freed is meant to be used, not to reduce the number of people in the team. He adds that he does not expect the gain to come from cutting headcount.
“make people more effective, not just more efficient”
He says he has wanted the team that processes renewals to be more proactive about customer health, and that before agentic capabilities he was not sure there was a way to get there. The company is now implementing an agent to take on renewal preparation and sentiment reading. He says the aim is to deliver the same experience to all customers while freeing the team to do more proactive outreach.
“this is now going to be our digital workforce that's going to be preparing all of the renewals”
Collins said the benefit goes beyond CSMs doing existing tasks faster. He said AI tries to connect data and create patterns, which can help solve problems the organization did not know it had. This is a forward-looking view and he did not describe a result from Dell.
“it's trying to connect and create all these patterns and find patterns that you didn't even know were there.”
said Jeanne DeWitt Grosser (Grit), Lauren Hughes (The Revenue Leadership Podcast), Jo Massie ([Un]Churned), Jim Richmond ([Un]Churned), Mark Wayland (The Science of Scaling)
Vercel grew from about 600 to just over 800 people in her roughly year and a half there, while revenue grew 'well north of triple digits'. She said she picks companies where revenue grows roughly exponentially while headcount grows roughly linearly, as at Google and Stripe. She attributes part of the difference to AI giving many roles multiple-times productivity.
“there are a bunch of things where like, you know, the equivalent human is getting NX productivity.”
Using Tangelo, which is tied to the role assigned in Workday, all scheduling for a new hire's three to five weeks of onboarding is generated automatically. Weekly knowledge-check tests run through AI, and managers get automatic alerts on progress. The early alert system creates a forcing function for a serious conversation with the manager about whether a hire will make it.
“all of your scheduling for your first, whether it's three to five weeks in onboarding is done now straight up through AI.”
Jo Massie said the customer voice team of four could no longer be afforded, so she asked the team to deliver the same outcomes with about 1.5 full-time employees. She said the team did amazing work with AI and exceeded her expectations completely. She said the program's output feeds the product team's roadmaps.
“So earlier this year I was like, we can't afford to have 4 full time employees doing this anymore.”
Richmond said he agreed that copilot-style applications give one person much broader depth of knowledge. From an operational leverage standpoint, his aim is efficiency, meaning people can cover many more accounts because they have a clear picture of what they need to do and tools to do it very efficiently.
“I want, I want people to be able to to, to cover a lot more accounts because they've got a very clear picture of exactly what they need to do”
Wayland notes that friends from Salesforce are now at Qualified and that many companies are building AI SDRs. He says he no longer needs the same SDR capacity because AI tools can do some of the work, which allows a different SDR-to-AE ratio. He frames the choice as an open question: take the change as savings, or use it to fund initiatives Box previously couldn't afford.
“I don't have to have the same SDR capacity that I did before because I can use AI tools. So then if I can have a different ratio between SDRs and AEs, do I take that as savings or do I take that to fund other initiatives that maybe I couldn't afford before?”
Teams under cost pressure or scaling efficiently bank AI gains as smaller headcount, while those focused on growth and value reinvest freed time into new work.
said Ross Simmonds (The Dave Gerhardt Show), Aviv Canaani (The Revenue Leadership Podcast)
Ross's example is podcast repurposing. Finding key moments used to take hours of listening before the clips were sent to an editor, and AI can now do much of that work. He says AI has created challenges, but he sees the opportunity as massive.
“We now have the ability to distribute great content, valuable stories across multiple platforms for a fraction of the cost that it used to take.”
Aviv says a new video used to need a script, a production company, and around $30,000 to $40,000, and now the team produces videos in days with tools such as Sora. He says the marketing team produced more than 100 campaign pieces in a month, which he could not have done without AI. He gave a Bob Ross Sora video on TikTok that got more than a million impressions as an example.
“it's going to cost you 30, $40,000”
said Alex Mashrabov (The Twenty Minute VC), Marc Ferrentino (Topline)
Mashrabov cites Higgsfield's open-sourced AI-generated movie project. In it, average prompt length was over 3,000 words and each scene used at least 10 image references to define characters, backgrounds and positioning. He compares video models to a modern rendering engine like Unreal or Unity, which he says cannot be directed through text alone.
“for 90 minutes of, let's say, TV quality content, it was over 100 hours of AI-generated content. So creative decisioning, picking the right piece, is still very important.”
Marc says that despite tools that make video creation look easy on social media, a video he made took him about 27 straight hours to look the way it did. He says it is not a one-shot process and still requires taste and energy. He expects video to keep thriving and wants easier ways for people without artistic ability to produce something that does not look like garbage.
“it still took me like 27 straight hours to get it to look the way it did.”
They differ on quality bar: volume repurposing and campaign content are cheap now, while polished creative output still needs heavy human effort.
What one named guest described doing or seeing. Each is a single account, not a point several operators agree on.
“it can really make you efficient in doing the wrong things. So if we don't have structure, it's just going to create a havoc.”
“what I thought I knew about our go -to -market process was proven wrong by something we learned recently.”
“if we're calling 85 % or 90 % of the funding rounds, two to six months in advance, and we're right 92%”
“when a. Customer has asked for something in the past this is the guidance that the CFO CRO and VP of CS have given in this type of example. We still don't use it for approvals”
“that is matching style guides, so terminology, tone of voice, localized, all these things at a fraction of the cost and turnaround time.”
“can we kind of support it in a way that's cost effective, right?”
Wade said these workflows look like traditional Zapier workflows with an LLM step inserted to handle unstructured data, make a decision, summarize, or categorize. LLMs do well with lots of context, so deterministic steps can feed precise context and target the model at a specific task. He said this gives higher reliability for the whole flow.
“But then in the middle of it, you're inserting either an LLM step or, you know, an agentic step that handles that handles some like unstructured data”
He said banks do not apply AI all the way through a workflow. Instead, they use a language model for a single interpretation at the right point, heavily guardrailed, while the rest of the process is rule-based. He presented this as the way banks get to an acceptable error rate.
“the actual implementations are almost 99% deterministic, and there's this one little percent that actually gets them to gain, to reason at the right moment in time.”
Asad Zaman challenged whether humans checking every AI-generated forecast formula saves any time. Keith said the work was a one-time implementation: Lightfield's head of finance reviewed every formula and some prompts were converted to code. After that the team makes only fine adjustments week over week. Keith acknowledged 'some might disagree' but said LLMs are not the right tool for math.
“We converted some things from prompts to real code. I don't want LLMs doing math. I want real code doing math.”
After a host noted that a champion can mean something different to each of five reps at the same company, Alex Bilmes said an LLM will fill the gap with definitions from what it was trained on on the internet. Endgame runs agents over a company's methodology language to build what it calls a semantic model, a dictionary of what terms mean for that organization.
“we actually have Agents that go look at your methodology language semantic definitions and actually build what we call a semantic model, which is a dictionary on what terms mean for each unique organization”
Alex Bilmes said companies that mandated creating many agents to measure AI adoption found that agents hitting different sources returned different answers. His example was five agents asked for a company's ARR giving five different responses. Automated prospecting also produced different messaging to the same persona each time.
“If you ask five agents what your ARR is as a company, you're gonna get five different responses.”
Alex Bilmes described a customer where a rep communicated a whole new set of SKUs and a new product offering, with pricing and packaging, that the company did not have. He used it to argue that individual AI answers need to be replaced by a centralized, company-approved source.
“We had a customer where a rep communicated a whole new set of SKUs. and a totally new product offering to the customer, because ChatGPT told them to.”
He defined selling time as the share of a week a rep spends face to face with a prospect or existing customer. He said he sees a lot of evidence that AI can push that to 60% or more. Holding ACV, close rate, sales cycle, territory and ICP constant, moving from 30% to 60% would algebraically double productivity.
“best in class has historically been about 30%. And I see a lot of evidence that today's AI can push that to 60 % plus”
Roberge says reps have historically spent 25 to 30 percent of their week selling because of time on call prep, meeting generation, pipeline reviews and CRM updates, and much of that can be streamlined by AI. He says he has seen early evidence that the best teams can reach about 75 percent, and he wants more teams to measure it. He asks boards to quantify AI-native claims.
“it has historically been best in class in the industry around 25 to 30 percent.”
Moss says users' default was to take a bad output and go do the analysis the old way elsewhere. Experity's message to operators was that they no longer need RevOps for the information, but they must teach the system why an output is wrong and what context is missing. He calls this a paradigm shift that was part of the process change.
“If the system gives me a bad output and I know that it's a bad output, I'm actually teaching it that it is a bad output”
Adrian said the agent reads a Slack channel for requests, pulls 60 days of customer call conversations, checks his calendar and Gmail, and saves a rewritten draft in Gmail. He built it in Claude Code with MCP connections to Slack, call data, Gmail and Google Calendar, plus a voice skill trained on emails he had already sent. It grades its own drafts, scoring lower when he edits them or gets no reply.
“right now it's actually gone close to 90% of the time I don't have to edit it.”
Kyle Norton says his team first had reps approve or edit what Momentum synced into Salesforce. The team compared the information reps entered with what Momentum captured against transcripts, and found Momentum more complete and more accurate. They then let Momentum write all the fields, and reps can edit them if they see an error.
“we just flipped it to like no, like Momentum's just going to write all the fields. We're never going to ask the reps.”
Kyle Norton says his team's AI scans every call and fills in all the fields. The team writes prompts and back-tests and backfills them on previous calls to see what happened when reps gave particular answers. He says they use prompts to tell whether a newly rolled-out pricing change is working.
“It scans every call and fills in all the fields and we can, you know, like write prompts and then back tests and backfill from previous calls”
Kyle Norton says his company's centralized approach emerged early, when a small group became obsessed with applying AI. They built things and deployed them to reps inside the surfaces they already used rather than asking everyone to learn new tools. He says the central owner need not be one person or an AI tsar and could sit in engineering with applied AI or in RevOps.
“we just built things, deployed it to the reps in the surfaces that they already use”
Kyle Norton says the popular advice is to give everyone Claude accounts and have them build, and he pushed AI adoption heavily about 12 months ago. In practice, though, the implementations that worked were highly centralized. A small group (the VP of BizOps and data, the VP of RevOps and Kyle) drove the work, and the VP of BizOps and data did most of the building. The tools were then deployed into the tools reps already used, such as Salesforce and Salesloft, rather than as new apps.
“what we really found in practice is that all implementations were all highly centralized.”
Moss says the fix was, first, access for everyone and, second, tools connected to the right systems and context on AgentCore. He notes that the agent is the easiest part to build once the framework, tools and context exist. He also hedges that being a regulated healthcare company probably helped limit rogue building.
“We only had a set group or a set number of people that had it. And what we learned was, A, it pissed a bunch of people off”
Rutten says GTMOS runs on enterprise authentication, so people log in with their own entitlements, and the Salesforce, research and email connections are made once at enterprise level, metered so no one can pull the full CRM at once. He says the biggest hurdle in building it was getting compliance sign-off for a non-regulated business that sells to regulated banks.
“we basically do that at a enterprise level so we have these tools under the hood fully connected fully guarded and metered”
Running his own informal evals for an app he is building (SaaStr Connect), the new Sonnet 5.5 used 42% more input tokens and 44% more output tokens. Quality improved and it passed more blind tests. He concluded that task-level cost is hard to predict.
“Input tokens, 42 % higher than before output tokens, 44 % higher.”
Having recently swapped models himself, Jason Lemkin said you have to requalify prompts and redo workflows, contrary to what 'the internet says'. He still called it much easier than swapping out a database.
“You have to qualify your prompt. You have to redo your workflows.”
Rory O'Driscoll said enterprises have so far defaulted to frontier models because people are lazy and bills were not astronomical. If OpenAI and Anthropic each reach about $70B in ARR, that $140B becomes a meaningful chunk of US corporate profits. He expects CFOs to push for cheaper models on non-frontier tasks, keeping Anthropic for the hardest problems, and to slowly take 'slugs of revenue' from the frontier labs.
“We need to shave three million off this token bill.”
Potter says organizations shifting to homegrown AI agents and Slack-based interfaces must also build how they maintain, version and iterate on those tools, and plan for a system crash at 3 a.m. He compares it to earlier cycles when companies built their own mail, documentation, procurement and data warehouses and later handed them to vendors, and says some companies are already pulling back.
“when you start to build your own engineering aspect, you also have to build the way that you're going to maintain those, the way that you're going to version those, the way that you're going to iterate upon those.”
Simon said that if a tool is used by an individual, building it yourself is awesome, but if more than a handful of people are using it, you really have to go with the vendor. For products sold to a customer base, traditional software engineering methodologies remain as important as ever.
“If it's an individual tool, building it yourself is awesome. If more than a handful of people are using it, you really have to go the vendor player.”
About three years earlier, Deltek weighed building and buying AI capabilities for support and ran both in parallel. Margo Martin said the tool they built worked and could have been rolled out, but they could buy basically the same thing without a team to keep it running as the models changed. Since then she said Deltek has mostly bought, while still playing around with building to some extent.
“we could buy basically the same thing. And then we didn't have to worry about keeping it up to date, right? The models were changing.”
Monika Saha describes a recent QBR with her go-to-market leaders, where each leader in her department brought one or at most two processes that are very painful for their teams. Processes were recorded on video so they could be shown rather than described, and the cross-functional group then watched each one. The group decided which processes to prioritize for solving with AI.
“Each leader in my department brought one or maximum two processes that are incredibly painful for their teams.”
Sam Jacobs suggests leaders might give people a framework with timed steps: the first 30 minutes listing activities and actions in the job, the next 30 minutes mapping those to things a computer or agent could handle, and then seeing how far they get trying to build it. He says if people get to that point, it is probably a lot more powerful and immediate for them. He offers it as a suggestion for leaders to give their teams.
“maybe give them a framework like let's you know the first 30 minutes you're gonna list activities and actions that you take in the course of your job and then the next 30 minutes you're gonna map those to things that you think a computer or an agent could possibly handle for you”
Asked about token budgets, he said the purpose must be specific, and at AIG it is the underwriting workflow. The aim is to get an underwriter as much perfect information as possible in a fraction of the time, then use more compute on models to shape the portfolio. He said the framework introduced at investor day a year earlier has advanced, and the next gains are in cycle time, data quality and getting information to underwriters faster.
“you need to be very specific on the business purpose or the functional purpose of what you're trying to deliver.”
Asad Zaman said the bar is something that really works, that you can trust to have done the job correctly, and that would change the company. His example was believing AI can run a sales cycle start to finish for anything under $20,000, 'maybe 50'. He said middle-of-the-pack models can't do these things effectively today and would add cost and risk. With the current top model, though, 'we're not that far', and the next frontier releases in about six months might do much more.
“think about what you'd be willing to spend a million dollars on in the next 12 months on AI.”
Sam Costello says buying processes now include AI committees that review all investments, and AI addenda that may be needed on existing contracts. As an illustration, he describes how a deal forecast in commit, with an MSA already in place, could slip if purchasing finds no AI addendum with two weeks left in the quarter and sends it into a legal process. He says the team triangulates with peers to find such pitfalls, though it is still difficult.
“there's now an AI committee that wants to review all the investments”
said Roberto Rivera (Impact Pricing), Keith Peiris (Topline), Tim Rutten (Topline, The Revenue Leadership Podcast), Kyle Norton (The Revenue Leadership Podcast), Jordan Crawford (The Revenue Leadership Podcast), Wade Foster (Topline)
He says that once AI has been trained with skills, business context and information, a price should come from a deterministic method of inputs, formulas and math. He argues computation is cheap, so the system can produce several perspectives that a team can evaluate. Whether a team accepts the output is still up to its own judgment.
“you don't let it when it comes to pricing right you don't let it guess you don't let it go and search the web and give me the answer”
Keith separates the deterministic numbers, which are handled in code, from the qualitative insight LLMs add. His example is looking past last quarter's closed count to ask how many deals nearly closed or slipped because of missing features or competitors.
“how many of them were sort of on the cusp of actually being able to close because of missing features, your competition and so forth.”
Asad Zaman challenged whether humans checking every AI-generated forecast formula saves any time. Keith said the work was a one-time implementation: Lightfield's head of finance reviewed every formula and some prompts were converted to code. After that the team makes only fine adjustments week over week. Keith acknowledged 'some might disagree' but said LLMs are not the right tool for math.
“We converted some things from prompts to real code. I don't want LLMs doing math. I want real code doing math.”
He said banks do not apply AI all the way through a workflow. Instead, they use a language model for a single interpretation at the right point, heavily guardrailed, while the rest of the process is rule-based. He presented this as the way banks get to an acceptable error rate.
“the actual implementations are almost 99% deterministic, and there's this one little percent that actually gets them to gain, to reason at the right moment in time.”
He says there is a misunderstanding that everything must go through an LLM. In GTMOS most of the system is deterministic code, with automation and workflows around certain activities, and LLM calls reached through an AI gateway only where needed. He frames the question as finding the right balance.
“I would say that almost almost 70 to 80% of GTMos is not LLM based. It's actually very deterministic”
Kyle's test is to define the business challenge, list the ways to solve it, and use AI when the work needs a large amount of data ingested and a non-deterministic judgment, such as pattern recognition. He said an agent at every step often leads to sloppy outputs.
“I think people too often it's a hammer looking for a nail and it's like what can I do with AI? And that's fundamentally sort of the wrong approach.”
He gave the example of deciding whether a prospect's tone is right to ask for a credit card, which he said AI should not do. He said the tools lack the capability for tone analysis. He said agents do well when tasks are split into deterministic choices and when they handle non-deterministic inputs such as a website.
“tone analysis is like something that, you know, the tools don't have the capability to do”
Wade said these workflows look like traditional Zapier workflows with an LLM step inserted to handle unstructured data, make a decision, summarize, or categorize. LLMs do well with lots of context, so deterministic steps can feed precise context and target the model at a specific task. He said this gives higher reliability for the whole flow.
“But then in the middle of it, you're inserting either an LLM step or, you know, an agentic step that handles that handles some like unstructured data”
said Justin Shriber (Topline), Tim Rutten (The Revenue Leadership Podcast), Alex Bilmes (Revenue Builders), Greg Casale (Revenue Builders), Kyle Norton (The Revenue Leadership Podcast, Topline)
He said the centralized model should create playgrounds for different functions, with constraints so that any question about revenue returns consistent data based on the user's role and security access. He also said that teams given free rein burned tokens quickly and were inefficient, while centralizing lets the company optimize spend against output.
“I think ultimately the answer is a centralized model that creates these playgrounds for different functions”
He described a dinner conversation with other operators where people had given every team access to agent tools, and in a meeting each group's data set, built with its own agents, did not match. He said the meeting became a discussion of where the data came from rather than what the company should do.
“none of the data matches”
Rutten says GTMOS runs on enterprise authentication, so people log in with their own entitlements, and the Salesforce, research and email connections are made once at enterprise level, metered so no one can pull the full CRM at once. He says the biggest hurdle in building it was getting compliance sign-off for a non-regulated business that sells to regulated banks.
“we basically do that at a enterprise level so we have these tools under the hood fully connected fully guarded and metered”
Alex Bilmes said that when each of a thousand reps has a different value proposition, sales process and way of describing value to an ICP, coordination and momentum suffer. He argued for combining customer data, methodology, enablement assets and messaging into one foundation that any human or agent queries, so answers align with corporate strategy.
“Centralization is a lot more important.”
Alex Bilmes described a customer where a rep communicated a whole new set of SKUs and a new product offering, with pricing and packaging, that the company did not have. He used it to argue that individual AI answers need to be replaced by a centralized, company-approved source.
“We had a customer where a rep communicated a whole new set of SKUs. and a totally new product offering to the customer, because ChatGPT told them to.”
Greg Casale says giving each rep an AI sourcing tool does not produce the value; instead, the team pulls ICP, TAM accounts, contacts, phone numbers and propensity signals into the back office. He says the prepared data is sent to SDRs through Sales Loft, which removes human sourcing errors such as the right person at the wrong company. He says that research which used to take hours a day is now already in place when a rep makes the call.
“instead of arming each rep with an AI based sourcing tool so they can get really smart about who they're calling that's not that's not the value”
Kyle Norton says his company's centralized approach emerged early, when a small group became obsessed with applying AI. They built things and deployed them to reps inside the surfaces they already used rather than asking everyone to learn new tools. He says the central owner need not be one person or an AI tsar and could sit in engineering with applied AI or in RevOps.
“we just built things, deployed it to the reps in the surfaces that they already use”
A host says decentralized AI gives the consumer benefits that individuals feel, such as emails written faster, meeting notes and decks built faster. He says it does not produce use cases that can be shown to move the business in a substantial way and are ready for production. He describes the output quality as depending on each person's taste.
“But you don't get like real business use cases, like the ones that you can say, ah, this move the needle for the business in a really substantial manner”
Kyle Norton says the popular advice is to give everyone Claude accounts and have them build, and he pushed AI adoption heavily about 12 months ago. In practice, though, the implementations that worked were highly centralized. A small group (the VP of BizOps and data, the VP of RevOps and Kyle) drove the work, and the VP of BizOps and data did most of the building. The tools were then deployed into the tools reps already used, such as Salesforce and Salesloft, rather than as new apps.
“what we really found in practice is that all implementations were all highly centralized.”
said Jacob Moller Korsgaard (Impact Pricing), Roberto Rivera (Impact Pricing), Mark Stiving (Impact Pricing), Sam Jacobs (Topline), Katie Bullard (Topline), Diane Wu ([Un]Churned) and 1 more
Danfoss has built a system that helps sales teams work smarter, has more advanced analytics, and is building its own pricing software with AI embedded to help users set the right prices. Jacob says AI is very good at using existing data and extrapolating from it, so humans must control it and add commercial judgment it cannot generate itself.
“we have quite good master data which is really fundamental to make AI work.”
He describes the scar tissue from implementing approaches that did not work, which lets experienced pricers recognise when an AI suggestion is unlikely to work. He says training AI takes time that not everyone has, and encourages the next generation to still learn what good looks like by making the mistakes experienced pricers made.
“And we know what good looks like, right?”
Mark Stiving says AI is wrong on pricing questions he knows the answers to, yet he believes it on unfamiliar topics such as a skin rash, which he calls dangerous. He says he tries to use AI as a first-pass answer and does not necessarily trust it.
“I know that it's wrong because I happen to know the answers.”
Responding to Mike, Sam said the primary feature of AI is persuasiveness and certainty, not necessarily accuracy. He said that is why subject-matter experts are needed to say when an output does not work or is not correct.
“The primary feature of AI is persuasiveness and is certainty. It is not necessarily accuracy.”
She compares AI-built models to an FP&A analyst's Excel model with one mistake, which can flow through a whole bookings model. She says a human still needs to check the work, and that AI makes human judgement on trade-offs more important rather than less.
“I worry about it if you're relying on it solely to give you the answer without the human review.”
Diane Wu says AI is a strong source of raw information and analysis but needs a CSM or CS leader with curiosity and technical fluency to prompt it well and ask the right discovery questions. She says this human judgment is what turns output into differentiated value. She frames the human role as curating and adding context, not replacing the CSM.
“when you use an AI tool, never ever take it for face value, right?”
Nick said he has seen teams damaged by using generative AI on top of complex models after they made mistakes by trusting what the bot said. Because Dreamdata's customers take its numbers to their boards and C-suites, he said the company approaches AI carefully so its output is accurate enough for decisions.
“I've seen teams wiped out utilizing generative AI on top of these complex models completely because they made mistakes by just trusting what the the the the bot saying.”
said Rob Potter (Revenue Builders), Snehal Nimje (Topline), Margo Martin ([Un]Churned), Joubin Mirzadegan (Grit), Jim Richmond ([Un]Churned)
Potter says organizations shifting to homegrown AI agents and Slack-based interfaces must also build how they maintain, version and iterate on those tools, and plan for a system crash at 3 a.m. He compares it to earlier cycles when companies built their own mail, documentation, procurement and data warehouses and later handed them to vendors, and says some companies are already pulling back.
“when you start to build your own engineering aspect, you also have to build the way that you're going to maintain those, the way that you're going to version those, the way that you're going to iterate upon those.”
Snehal said organisations that try to keep updating models, workflows and MCPs themselves may spend three to five times more resources. He said a proof of concept or MVP might work for two or three months but then become outdated. He said customers who tried building this way later came back looking for a solution to the problem.
“you're spending your your spending 3x or 5x more resources, right? And you will be outdated in six months or so.”
About three years earlier, Deltek weighed building and buying AI capabilities for support and ran both in parallel. Margo Martin said the tool they built worked and could have been rolled out, but they could buy basically the same thing without a team to keep it running as the models changed. Since then she said Deltek has mostly bought, while still playing around with building to some extent.
“we could buy basically the same thing. And then we didn't have to worry about keeping it up to date, right? The models were changing.”
His list: an AI model building the data model and the pricing and rules engine from scratch with good opinions, and an agent architecture on top. Beyond that, the tool must serve IT, deal desk, RevOps, sales and finance workflows, integrate with billing and CLM so amendments and renewals are right, and handle the legal risk of sending wrong quotes. It must also be load-tested for thousands of salespeople, maintained as people leave, and reworked when the CEO wants a new token-based pricing model in two months. He adds that a model drawing on prior CPQ tools will reproduce 'some tool designed for IT and RevOps that sales absolutely hates.'
“CPQ is unique because it touches IT, deals desk, revops, sales, finance, minimally, just to get from a quote to a deal approved to the customer.”
He argued that a company building a highly specialized piece of software with general AI tools will not get something better, cheaper or easier to maintain than an off-the-shelf product, because the risk and benefit are not in line. He expects specialized AI applications to win in customer success and support, where problems are smaller, tightly controlled and have data to train models.
“a lot of the big winners that are going to be out there over the next few years are going to be AI application companies”
said Dan Lee (The Science of Scaling), Ann Davis (Revenue Builders), Mark Roberge (The Science of Scaling), Michelle Bove (Revenue Builders), Joubin Mirzadegan (Grit), Frederic Kerrest (The Science of Scaling)
Dan Lee said models are good at coding because code is fully observable and can be compiled and tested. Sales is the art of influencing a buying decision across several people, and you may not know how a conversation went until months later. He said sales is also adversarial, multiplayer and under-constrained, which is why he thinks it needs human taste and judgment.
“I don't know how that uh you know, how that sales conversation went until 3 months later if they signed the contract or not.”
Ann says enterprise deals involve a lot of interconnected people at the buying company, and the seller is responsible for getting them all aligned. She says that is unlikely to change at the enterprise level, while a $5,000 deal can be automated.
“When you're talking enterprise, I'm talking $500 ,000 plus type deals.”
He agrees with Christopher that taste, which he calls creativity, will remain human, citing songwriters, film, books, art and sports. He has recently developed conviction about human-in-the-loop roles, comparing them to pilots who remain even though planes can fly themselves. On a million-dollar software purchase, he wants to shake a human's hand to confirm his needs are understood and will be delivered.
“And if I buy a million dollars of software from someone, I'd like to shake a human's hand to make sure that they have blessed my needs and the fact that they will actually deliver.”
Bove says AI can shorten the work for reps who run many small transactions, but reps on strategic accounts need to be on site with executives, champions and procurement. She gives the example of a rep who spends most days with customers in person and closes large deals quarter after quarter, and says remote-only competitors may be at a disadvantage.
“But if you're doing more strategic selling, you have to be in person.”
Joubin Mirzadegan cited a tweet from someone on his team saying we wouldn't tell engineers they no longer need to code because Claude Code exists. He said the same applies to sales, which does not stop, and the conversation concluded that people are still needed to run things through to completion.
“It's the same thing with sales. Like, you know, like, this doesn't stop selling.”
Kerrest says agents won't be buying seven-figure deals from agents anytime soon. At that deal size it is people buying software from people, which he says is why founders need to understand and build those relationships, and why enterprise sales as a profession is not going anywhere.
“So enterprise software, where you're trying to do six -figure, seven -figure, eight -figure ARR deals, will, for now and for the rest of my lifetime, be done between people.”
Schuck said AI to AI to AI makes sense in transactional moments, and AI to human also works there. He said that as deals become bigger and less transactional, they are further from being replaced by AI. He was responding to a question about whether AEs and demos would be replaced as well as SDRs.
“I think when things become like bigger and less transactional, that that's a further away from being replaced.”
Rosenthal described a hypothetical future in which a rep's AI assistant prepares briefings on a company and its people, listens in on calls with hints, and updates the CRM with next steps, follow-up emails and decks. She said the assistant would also handle contracts, red-lines, RFPs, questionnaires and CRM entry. She said she thinks it is unlikely that people will still make big purchases without looking somebody in the eye and having a conversation.
“all of the minutia of sales, all of the mundane administrative tasks, all of the managing contracts. and tracking down legal documents, getting them red -lined answering RFPs, answering questionnaires, filling out your sales where CRM, all of that stuff will be handled by your assistant, your AI assistant.”
said Ian Tickle (The Revenue Leadership Podcast), Simon Farthing ([Un]Churned), Saumyo Mukherjee ([Un]Churned), Amanda Moran ([Un]Churned), Daniel Simon (Revenue Builders), Alex Bilmes (Revenue Builders)
Referring to a previous episode about building AI agents, Ian raises the question of how companies will manage all the agents they create, and who they will hire to manage AI. He asks where discipline and control sit if every department builds an agent to patch its own problem.
“You imagine if every department creates its own AI agent to fix a problem they've got while fixing the core systems That's a lot And then where is the discipline and control?”
Bloomreach tested about 60 AI tools across its go-to-market teams over the last 18 months, including Gong, Zoom, Granola, Glean and several note takers. Simon said people had too many choices, nobody was consistent, and the company could not measure anything because everyone was doing something different.
“I think what I've learned from that is tool proliferation is its own problem.”
Because the business systems team supports functions across marketing, sales, services and success, it sees where AI tools overlap or conflict. Saumyo says the team has to make sure there is proper guardrails, no security gaps, no overlapping functionality and no confusion about which tool to use. He says that a person not knowing what to use would ultimately become the team's problem.
“So we are a point of convergence if you will and it's our responsibility to make sure that we have proper guardrails there”
Saumyo's example is a sales rep who vibe codes an automation for a process that needs a VP approval, which works for that rep. He says there are at least three other tools already doing the same automation, maybe better. This is why the business systems team has to manage overlap across the company.
“There are at least three other tools that do that same automation maybe better”
Everyone at Darktrace is testing agents individually, with cross-team sharing about whether work can be done more efficiently or smartly. A new agentic Tiger team is launching to coordinate agent development, surface needs, connect agents and orchestrate work. Amanda Moran said that coordination will be critical as the company moves into its next phase of working internally and with customers.
“We also have an AI or an agentic Tiger team that is launching to coordinate the way that we're developing agents across the business.”
Daniel said companies have many agents built by different teams, some of which he did not know existed. He said Glean can break agents down by function such as sales and show how often each is used, and that helping with this problem is a big opportunity in the market.
“I think there needs to be a management layer on top of all of those agents to better identify, like what are the best agents to use?”
Alex Bilmes said companies that mandated creating many agents to measure AI adoption found that agents hitting different sources returned different answers. His example was five agents asked for a company's ARR giving five different responses. Automated prospecting also produced different messaging to the same persona each time.
“If you ask five agents what your ARR is as a company, you're gonna get five different responses.”
said Grant Clarke ([Un]Churned), Jonathan Moss (The Revenue Leadership Podcast), Alex Bilmes (Revenue Builders), Adrian Rosenkranz (The Revenue Leadership Podcast), Brad Casemore ([Un]Churned), Pablo Dominguez (Topline)
Grant says many companies are struggling to deploy AI agents across technical services, early funnel, sales and renewals, and that it is 'difficult enough just to get the process to work with a human.' Atlas aims to build the frameworks, standards and loop process to train the agent faster because humans are deployed with it. It offers this as a service so the client company does not carry the burden.
“It's difficult enough just to get the process to work with a human.”
Moss says users' default was to take a bad output and go do the analysis the old way elsewhere. Experity's message to operators was that they no longer need RevOps for the information, but they must teach the system why an output is wrong and what context is missing. He calls this a paradigm shift that was part of the process change.
“If the system gives me a bad output and I know that it's a bad output, I'm actually teaching it that it is a bad output”
Alex Bilmes suggested the hosts consider agents rather than humans as the primary users of enablement and training, and asked what agent enablement, training and performance management look like. John Kaplan added that people should learn to orchestrate agents within their company's guardrails, with humans always involved.
“So there's a new concept, guys, that you might want to consider, which is agent enablement.”
He said each run of the email agent scores itself, with 100% meaning he didn't touch the draft and got a response. When a task is harder than it should be, he asks the agent to review the conversation, act on the actionable feedback and rewrite its approach. He said this has been a lifesaver.
“Time for a mini retrospective.”
Adrian said the first version of an agent is rarely the problem, and that keeping it working on day 20, 50 and 100 is. He said he had to change his habit so that each bad output leads to an update of the underlying skill and context, rather than retrying or dropping the output.
“the day one is never the problem. Making sure things are like working on day 20, 50, 100. That's the hard part.”
He says agents will make mistakes and should not be expected to be perfect when brought in, so the team has to be ready to train, coach and give feedback to improve them. He describes managing digital capacity in the same way the company manages human capacity.
“You have to then manage your digital capacity the same way.”
He said each agent is given its function, its input, the expected data structure, and the data sources it can check, much like training a person. He said the human is needed not because the agents hallucinate, but because they sometimes cannot do or find things and need guidance.
“What I've learned is you have to treat them like humans.”
said Christopher O'Donnell (The Science of Scaling), Kyle Norton (The Revenue Leadership Podcast), Jonathan Moss (The Revenue Leadership Podcast), Lauren Hughes (The Revenue Leadership Podcast), Brady Bluhm ([Un]Churned), Jordan Crawford (The Revenue Leadership Podcast)
Christopher describes skills as small training manuals or slash commands that let the AI do something perfectly every time, and calls them settled. Agents have job descriptions, a view of the world and opinions about what matters, and you need more than one. How skills get automated (loops, schedules, triggers) is still undecided and will have many vendors. The fourth layer is the data source, which he says should be the actual customer truth down to the word said, who said it and when.
“I think there are a couple parts of it that are really decided and that is the idea of an agent and the idea of skills.”
Norton is using a Grokbot setup agent to walk him through porting his OpenClaw instance. Most of it lives in the GitHub skill files, while items such as crons built inside OpenClaw's virtual machine have to be rebuilt.
“everything I build. Build is always instructed right Back to the GitHub repo with these skill files. So I can. I can use multiple experiences with the same tools and skills.”
His agents are grouped into teams: engineering, product, sales, marketing, customer success, RevOps, enablement, a GTM advisor panel (which includes a Kyle Norton persona) and a personal board of directors. A skill is a repeatable workflow, such as how to pull and analyze Google Analytics, and is kept out of individual agents so several agents can share it. A smart routing agent picks which agents to use based on his request.
“a skill file is kind of a repeatable process”
The whole company has Claude. Revenue effectiveness centrally manages the enablement skills, which hold the personas and connect to Tangelo onboarding results. For a launch, a skill produces the project plan a human used to write, lists what is needed from product, PMM and marketing, and generates PPTs, podcast-style recordings, Synthesia videos and AI role plays. Enablement staff, now called revenue readiness professionals, review outputs, and results are fed back to update the skills.
“Yeah, so the skills produce the project plan, which a human used to produce”
Brady Bluhm says that whenever he does any task, he first asks whether it can be done with AI. After getting a usable output, which he says takes curating context and multiple rounds of feedback, he asks whether he will need to do the task again. If so, he considers training an agent or skill so the next run is faster.
“I ask, can I be doing this with AI?”
Kyle used a meta prompt to ask a model to turn his writing into a skill, then gave that prompt all his writing. The output described his tone, patterns and structure in detail. He now has separate skills for newsletters, LinkedIn and email, and the email skill says never say hello and only write one line.
“I had this 150-page Google Doc of like all the LinkedIn posts I basically ever written. And I just dumped this into a model.”
Jordan said a client wrote out the steps to launch a campaign and he gave them to Claude as a markdown file. Each time Claude starts a campaign task, it loads these steps as the plan. Changes are made in the code repository and pulled down from GitHub.
“every time we launch a campaign we should do these 47 steps”
said Jonathan Moss (The Revenue Leadership Podcast), Ian Tickle (The Revenue Leadership Podcast), Amanda Moran ([Un]Churned), Pradeep Raman ([Un]Churned), Kyle Norton (Topline), Carsten Schütz ([Un]Churned) and 2 more
Starting with agents and interface, he kept hitting the same problems: agents lacked context, couldn't work across multiple systems and tools, and couldn't do what he needed. Solving those pushed him all the way down to the data, which he calls a lesson learned.
“I started at the agent and interface, and every time I started working to build the system, I kept running into things. Why does it not have context?”
Ian says AI is phenomenal only if the systems and knowledge management are in place. If systems are not connected, AI ends up acting as a shadow RevOps function. He says the organization then loses control over what it is trying to do and over its reporting and information.
“If your systems aren't talking to each other if they are not connected then what you tend to find is that AI's actually doing a shadow revops org”
Amanda Moran listed the data to connect: product usage, contract systems, all customer communications, community activity and learning center activity. She said LinkedIn had built a scaled digital program over many years but still had things that were not connected in the background. She described connecting this data as the manual first step toward using agents.
“the sort of manual, in a way, first step, I think is connecting all of those data components.”
Pradeep argues that AI transformation is only as good as the location of the customer's data and applications. If data and apps are in legacy environments, he says it is very hard to maximize the full potential of AI transformation. This is why modernization is a top priority alongside agentic AI.
“If your data and apps is sitting in legacy environments, it's very hard to, to really maximize the, the full potential of AI transformation.”
Kyle Norton recommends that every company start with good first-party and third-party data. He says without it the models cannot be given what they need to make smart decisions or produce good outputs. He describes this as the foundation that the rest of his AI work at Owner.com is built on.
“The recommendation I give to everybody is start with data. You have to start with good first party and third party data.”
Carsten tells listeners who are customers to make a plan for the baseline before adopting AI. He lists cleaning up data and comparing the technology stack and data to prepare for AI and agent capabilities.
“So how do you clean up your data?”
Collins said the foundation work of standardizing data sources, KPIs, triggers and playbooks is not fun, but it becomes the raw material for AI. He said AI is then layered over the top to bring data points together more predictably and proactively. Dell's scale lets it do that groundwork well.
“let's get standardized triggers, let's get standardized playbooks. That all becomes raw material to work with AI.”
He said people sometimes plug AI in and only then realize their data is bad. A co-host then described an Insight portfolio company whose business is fixing bad enterprise data, and argued for simplicity, and Pablo agreed.
“And now all you've done is automate trash.”
Schuck said service AI tools such as Zendesk and Intercom work well because they use first-party data such as knowledge bases and tickets, which is complete and accurate. He said go-to-market AI needs third-party data as well, and that CRM data is not accurate enough to power it. He said companies have barely mastered their CRM for existing functions such as renewal tracking.
“And nobody has actually like spent a lot of time to master their data in their CRM to be ready for a universe of AI.”
said Mark Thurmond (Revenue Builders), Mark Roberge ([Un]Churned, Topline), Sam Jacobs (Topline), Kellie Snyder ([Un]Churned), Frederic Kerrest (The Science of Scaling)
Mark said Tenable wants its sellers doing seven or eight face-to-face sales calls and 10 to 15 Zoom calls, and has very high activity expectations. He cited reps freeing 15, 20 or 25 percent of their week from non-customer work by using AI as an enhancement and co-pilot. He said he expects these productivity gains to continue for years, while describing the company as only in the early stages.
“Being able to free up 15, 20, 25 % of their week by using AI is phenomenal.”
Roberge describes the first phase as eliminating admin work to increase selling time, which he says is happening now. He expects a second phase where agents are sellers and a third where agents are buyers. In a fourth phase he expects functional divisions such as sales and marketing to lose their shape because they were designed around human limits.
“The first phase is just completely eliminating the admin work and just increase in selling time”
Roberge says reps have historically spent 25 to 30 percent of their week selling because of time on call prep, meeting generation, pipeline reviews and CRM updates, and much of that can be streamlined by AI. He says he has seen early evidence that the best teams can reach about 75 percent, and he wants more teams to measure it. He asks boards to quantify AI-native claims.
“it has historically been best in class in the industry around 25 to 30 percent.”
He listed pre-call prep, post-call follow-up and CRM updates, saying pre-call prep and after-call follow-up could become much quicker, maybe doubled. Roberge said the framing made sense to him and that these administrative gains feed into selling time.
“So what I hear you saying is the AI use case and go to market is the administrative work.”
He defined selling time as the share of a week a rep spends face to face with a prospect or existing customer. He said he sees a lot of evidence that AI can push that to 60% or more. Holding ACV, close rate, sales cycle, territory and ICP constant, moving from 30% to 60% would algebraically double productivity.
“best in class has historically been about 30%. And I see a lot of evidence that today's AI can push that to 60 % plus”
She described using AI to handle the data-entry tasks that take time, such as capturing meeting outcomes from tools like Gong and noting whether customers are happy. In her description, a CSM would then get a prioritized list of the next things to discuss with the customer across upcoming meetings. She framed this as an approach that is developing over time.
“Did this meeting happen? What did the customers say? Are they happy? Are they not happy? All of that, like taking that all off the table”
Kerrest says AI can help salespeople focus, qualify leads faster and flag companies in a lead set that should stand out but aren't obvious. He sees prep work, meeting notes, filling templates and internal up-and-down communication as busywork that can be automated, without AI replacing salespeople.
“how much work did we used to do as sales professionals, just prepping things and writing our meeting notes and filling out the templates and making sure that we communicated up and down our internal organization. And a lot of that can be automated.”
said Jeanne DeWitt Grosser (Grit), Rebecca Nerad ([Un]Churned), Bob London ([Un]Churned), Jason Forget (Revenue Builders), Kyle Norton (Topline)
Grosser described analysing call transcripts for patterns as a way to get product and go-to-market to agree on what is and isn't working. Objections that can't be handled because the product lacks something are automatically submitted as product asks in Vercel's internal GTM feedback tool.
“we proactively extract objections from calls and score the degree to which we've handled them.”
Nerad said the tool gives individual CSMs specific feedback on their conversations and gives feedback to their managers. She receives a weekly trend report that calls out specific moments from customer meetings and themes across a product set. She said the sentiment readings lead CSMs to ask whether a customer's frustration should change the health score or how they respond.
“To me as a leader, I get a weekly report of trends and it will call out specific things from customer meetings that I'm also able to look at trends across a particular product set or a theme.”
London built a custom GPT trained extensively on his 'radically authentic discovery' methodology to assess calls. It produces a coaching scorecard with a timeline in three columns: what the customer said, what the rep said next (for example, did they dig in or start a new thread), and a coaching point on what to say next time. He says doing this manually took him about two hours per transcript.
“it does would take me two hours per transcript and I'm not sure that's the work I actually want to do anyway.”
Jason said Cockroach Labs built a tool called Deal Pulse, which is an MCP server connected to its data sources and built on its sales process and qualification. He said it can infer an economic buyer from a conversation even when the buyer is not recorded in the CRM. He described this as a way for managers to spend less time checking process and more time coaching.
“I can get inference around the economic buyer and test that, which is tremendously powerful”
Kyle Norton says his team's AI scans every call and fills in all the fields. The team writes prompts and back-tests and backfills them on previous calls to see what happened when reps gave particular answers. He says they use prompts to tell whether a newly rolled-out pricing change is working.
“It scans every call and fills in all the fields and we can, you know, like write prompts and then back tests and backfill from previous calls”
said Rory O'Driscoll (The Twenty Minute VC), Dev Ittycheria (The Twenty Minute VC), Jack Altman (The Twenty Minute VC), Jason Lemkin (The Twenty Minute VC), Tim Rutten (Topline), Matthew Kropp ([Un]Churned)
Rory O'Driscoll said enterprises have so far defaulted to frontier models because people are lazy and bills were not astronomical. If OpenAI and Anthropic each reach about $70B in ARR, that $140B becomes a meaningful chunk of US corporate profits. He expects CFOs to push for cheaper models on non-frontier tasks, keeping Anthropic for the hardest problems, and to slowly take 'slugs of revenue' from the frontier labs.
“We need to shave three million off this token bill.”
Dev Ittycheria said a model within about six months of the frontier that covers 90% of use cases would be attractive, especially for reasoning workloads like coding and agents. He said he didn't see why enterprises would not want to talk to such vendors immediately.
“this also reinforces the point that you don't need the frontier level intelligence for every workload.”
His examples: a tax return is either filed correctly or not, and a hammer only needs to drive the nail. He expects more open-source usage and strong inference companies as a result. He added that the labs are structurally cost-advantaged in compute, users and ways to subsidise, so open source will be a big part of the market but will not dominate.
“Once you have filed it correctly, throwing more intelligence at that problem, doesn't do you any good.”
In a mock investment-committee pitch for an AI model startup seeking a $10B valuation a week after its seed (rendered as 'Jeff' in the transcript), he said it already has 17% of OpenRouter traffic and 20% through Vercel's router, at one-seventieth of the price and 100x faster. He said that even with newer model versions getting cheaper, spending at current costs 10-12 hours a day is unsustainable. He acknowledged it may not be the winner next year.
“These AI costs are unsustainable. It doesn't matter if Sonnet 5 .5 and the latest Opus is cheaper. It is unsustainable to spend these costs 10, 12 hours a day”
He said there is already sufficient intelligence for banking and that the labs are competing like electricity providers, with models becoming a commodity. He argued that banks need smaller, more efficient models, better infrastructure, and a standardized architecture so every bank can be governed properly. He added that open-weight models mean this will not stay within one continent.
“we don't need more intelligent models. We need smaller models, more efficient models.”
He said a bank should use frontier labs for some workloads, but most daily workloads are repetitive and deterministic because banks follow checkpoints and policies rather than reasoning. He suggested a small language model trained on financial conversations or transactions could do better than the top model at a far lower cost. He described the top model as too expensive for this work.
“You don't need Mythos. Trust me, you really don't.”
This quote could not be matched to the transcript. Treat it as a paraphrase.
Kropp said he would not be surprised if future specialist models charged around $10,000 per million tokens instead of $25, a hypothetical figure he gave. He said he would use the expensive model for drug discovery and cheaper models for code or planning a vacation.
“maybe they're charging, you know, $10,000 for a million tokens instead of $25, right?”
said Ian Tickle (The Revenue Leadership Podcast), Parag Agarwal (Grit), Chael Banks ([Un]Churned), Brad Casemore ([Un]Churned), Jared Collins ([Un]Churned)
Ian says being quicker is not the same as being more efficient or effective, and gains that aren't redeployed wash out. His example is automating reporting so that the analysts who build reports can work on programs or ICP and TAM analysis instead. He frames the goal as using tools to move people onto higher-value work, not just to make jobs easier.
“what do you do in that efficiency to make value out of it? Otherwise there's like a net-net, it washes out at the back end. We're just being quicker but we're not necessarily more efficient or effective”
In what he calls a semi-hypothetical example, a PE firm that once used gut feel to narrow buyout targets to four or five could use Parallel and models to run near-exhaustive 'simulations' across criteria before choosing. Because capital and people limit how many deals the firm can do, picking better creates enough value to justify heavy token spend. He calls this the next year of growth: 'work that was not happening yesterday.'
“It's going back to what does it take to make it 1000 x, right? It is work that was not happening yesterday.”
His examples: backend engineers doing more frontend work, frontend engineers changing APIs, and people using agents to run security reviews or push optimisations they would otherwise never have done. He sees true value in people taking more end-to-end ownership and pushing their own learning curves.
“it's actually people expanding their span of influence and ownership on the product. That's where I think there is true business value being generated in my mind.”
Chael says Okta uses AI to remove low-value tasks so people can have higher-level conversations, and that the efficiency gains are in service of effectiveness. He says the time freed is meant to be used, not to reduce the number of people in the team. He adds that he does not expect the gain to come from cutting headcount.
“make people more effective, not just more efficient”
He says he has wanted the team that processes renewals to be more proactive about customer health, and that before agentic capabilities he was not sure there was a way to get there. The company is now implementing an agent to take on renewal preparation and sentiment reading. He says the aim is to deliver the same experience to all customers while freeing the team to do more proactive outreach.
“this is now going to be our digital workforce that's going to be preparing all of the renewals”
Collins said the benefit goes beyond CSMs doing existing tasks faster. He said AI tries to connect data and create patterns, which can help solve problems the organization did not know it had. This is a forward-looking view and he did not describe a result from Dell.
“it's trying to connect and create all these patterns and find patterns that you didn't even know were there.”
said Manny Medina (Topline), Ramin Heydari (Topline), Jonathan Moss (The Revenue Leadership Podcast), Mark Roberge (Topline), Pablo Dominguez (Topline)
A host asked whether the clearest form of reimagining is to sell the outcome rather than the form-filling work, citing how factories were redesigned around electricity. Manny agreed. Another host added that chat interfaces should give way to software that already knows what to do.
“At the end of the day, the process exists for a reason. And that reason is the outcome. And all the middle steps are all negotiable.”
Manny says software vendors used to treat professional services as taboo. He is excited by a new breed of companies that sell software together with transformation of the customer's operations, rather than layering AI onto existing workflows.
“there's this whole new breed of companies who are like, who are selling software with transformation.”
Manny thinks experimental spend described last year better than this year. Now buyers know they will spend on AI for a given body of work, pick a vendor and commit for longer. He puts enterprise penetration at roughly 1%, because most companies are still running old workflows faster with AI rather than redesigning the role around agents.
“I think that was the case maybe last year. I think this year people are sort of like locking down into longer commits”
Heydari says many business tools are just adding "a little bit of AI spice," and their makers believe that makes them working AI systems. He argues it requires much more than that: the business has to rethink how it operates, with things as much as possible under one roof so decisions can be made quickly on current data.
“so many tools in businesses are just adding a little bit of AI spice with and they think this is the AI system that is working, but it requires a lot more than that”
His sequence: start with the business outcome or constraint rather than an AI use case; map the current workflow, including handoffs, exceptions, required data, where context drops, and which meetings exist only because information or decisions are locked up; redesign by deciding what to remove, automate or augment; deploy with change management and measurement; then run a learning loop of user feedback and tweaks before scaling. He compares this to deploying a product.
“AI is not something that you add into an existing process because then all you're typically doing is you're just adding one extra step or one extra layer, but you're not really solving the business outcome or the root problem.”
Mark used the Web 1.0 analogy, when the internet was seen as putting a brochure online and people could not yet conceive of user-generated content or companies like Uber and Twilio. He said he is not bullish on the survivability of many companies funded today because they focus on workflow streamlining. He said AI 2.0 will be about workflow reinvention, which is hard to conceptualize.
“I feel like AI 1.0 is about current workflow streamlining.”
The team member who led the work was asked to reimagine the whole diligence process from an AI lens. Pablo said it took about nine months to work out how to do certain things, and that the team had to learn it along the way.
“What I see people doing today is putting AI on top of existing processes”
Mark Bersh says he thinks the first phase of AI at the application layer has been iterative. He contrasts that with disruptive use cases that remove around 80% of a task and rethink it, which he says can make the incumbent technology irrelevant and give an attacker an edge.
“You know, I think like this first phase of AI at the application layer has been very iterative.”
said Keith Peiris (Topline), Jaleh Rezaei (Topline), Pradeep Raman ([Un]Churned), Jordan Crawford (The Revenue Leadership Podcast)
Asked how a probabilistic technology can support forecasting, Keith said Lightfield assumes models will stay non-deterministic. For most growth-stage customers, Lightfield does not tell them to leave everything on automatic: important fields are set so reps approve Lightfield's suggestions, which keeps them auditable. For processes like pricing, discounting, qualification and account research, he says the models need a lot of context to be reliable.
“If you don't have auditability of your fields, you'll never believe your forecast.”
Jaleh said Mutiny makes approval mandatory so reps must review and edit content rather than sending AI output directly. She said she believes reps will have better outcomes if they stay involved in the process. She raised this in the context of concerns that AI note-takers may make salespeople less present.
“we're really big on, you have to approve it before it gets sent to the customer.”
Pradeep says his team has built custom agents for specific process flows. An analyst or intake agent handles processing of incoming items. An architect agent reviews the data, including recordings and transcripts of customer calls the team was part of, and recommends how to address the customer's need. Pradeep says the recommendations go through human review and validation.
“we have an analyst agent or an intake agent that does the processing and architect agent that will review the data”
He said he does not trust the tool enough to vibe code writes to his CRM. He cited Jason Lemkin's production database being deleted after someone gave the wrong access token. He said his business is about deploying, not destroying, so he designs his work around structured outputs.
“I don't trust the tool enough to like vibe code and like write to your CRM”
said Jonathan Moss (The Revenue Leadership Podcast), Jared Collins ([Un]Churned), Guy Galon ([Un]Churned), Josh Schachter ([Un]Churned)
Moss says the fix was, first, access for everyone and, second, tools connected to the right systems and context on AgentCore. He notes that the agent is the easiest part to build once the framework, tools and context exist. He also hedges that being a regulated healthcare company probably helped limit rogue building.
“We only had a set group or a set number of people that had it. And what we learned was, A, it pissed a bunch of people off”
Collins said he has encouraged his team to experiment, asking what repetitive day-to-day tasks could be turned into a one-click action, and how AI might surface insights from a customer contract. He said he is only doing some Excel vibe coding so far, and that individuals have already produced useful results in a few minutes.
“It doesn't have to be these massive kind of enterprise wide kind of initiatives to get that AI efficiency. It starts with just day-to-day work.”
He says the team should not set big targets but should solve a small use case first and see that it works. He encourages this because he wants Obrela's AI next steps to come from his team, not because he tells them to adopt them.
“So don't try to set big targets. Try to solve a small use case first and see that it works for us.”
He says every organisation has people, perhaps an ops person or a CSM, who are tinkering with AI on their own. He advises making them the poster child of the organisation, encouraging them and feeding them the resources they need. He says you have to promote AI adoption when it comes organically from those people. Kristi agreed.
“you can't always force it. You have to really kind of like promote it when it's coming organically from from those.”
said Ann Davis (Revenue Builders), Bryan Murphy (Topline), John Kaplan (Revenue Builders), Asad Zaman (Topline)
The investor wasn't using specialized tools, just basic workflows while switching between Claude and ChatGPT. He now spends 25% of his former time on the same work, freeing 75% for other things. The speaker presented it as an example of how well AI works where it fits.
“He's like, my job is now 25% of what it used to be.”
Ann describes a financial analyst whose job is to research new markets and companies, and says the job can be done in about 30 minutes rather than a full week. She presents this as an example of the productivity gains from analysing data through an AI interface.
“They can do their whole job. of a full week as research in probably 30 minutes.”
He described this as AI with a human in the loop, which makes Smartling's translators vastly more productive. He compared it to developers using AI to write code, where the output rarely goes straight to production. He said the productivity figure is still climbing.
“It's like 8 ,000, 10 ,000 words a day and climbing”
Kaplan says he dictates voice notes while walking, riding a bike or at the gym, for pieces like his Monday motivations. Output that used to take him hours to write now takes probably less than 20 minutes. He says the notes need to be useful and purposeful rather than just vomiting up ideas.
“Those things would have taken me hours to write in the past. They take probably less than 20 minutes now from me which is fantastic.”
Asad Zaman says his recruitment company, Sales Talent Agency, spends an hour or two interviewing each candidate and then two to three hours writing up their profile. Almost within months of AI gaining momentum, the document creation dropped to about an hour and then to 45 minutes. He says early, easy wins like this made it easy for the company to believe in AI.
“that became an hour. And then that became that became 45, like the document creation portion of that.”
said Ross Simmonds (The Dave Gerhardt Show), Joubin Mirzadegan (Grit), Dave Gerhardt (The Dave Gerhardt Show), Sam Jacobs (Topline), Darren McKee ([Un]Churned)
Ross says low-quality AI content is not limited to Reddit but fills X, Threads, LinkedIn and Quora. He argues that brands that put authentic, valuable human content on these crowded platforms can win.
“That creates an opportunity If you can create authentic, valuable human content and then put it on these platforms, you can win.”
He described reps running meeting notes from Granola through Claude and sending the output to the team. He also said every AI-made deck now looks the same. He told his team that if he doesn't want to read these, customers definitely don't. Grosser added her own complaint: AI-written call prep docs instruct her in an unnatural tone, such as telling her not to sell on an exec-to-exec call.
“your job is to synthesize what actually matters.”
Dave briefed Claude with a brief, examples and screenshots, had it build the outline, and drafted from it, but the draft had no soul. He says he then asked Harry Dry for help, and Harry told him to put Claude away and write the talk himself.
“Time to put away Claude and get into your own head.”
Sam Jacobs says the agent drafted posts and replies that sounded good but had an artificial structure and rhythm. He says he turned it off, concluding that one authentic response is better than a hundred fake ones. He notes that the agent's running costs were also a concern.
“I realized the main thing it was doing was auto responding to every comment.”
He says he tells his trainees not to send a templated message to all prospects, but to study each company, how it makes money and its revenue, then write a tailored pitch. He is critical of sending that pitch through an agentic LinkedIn automation tool and pressing go. He says LinkedIn does not allow much of that automation, and that a human-written approach stands out in the feed.
“the problem with that is we've outsourced a lot of the actual effort that we do to build relationships.”
said Jeanne DeWitt Grosser (Grit), Lauren Hughes (The Revenue Leadership Podcast), Jo Massie ([Un]Churned), Jim Richmond ([Un]Churned), Mark Wayland (The Science of Scaling)
Vercel grew from about 600 to just over 800 people in her roughly year and a half there, while revenue grew 'well north of triple digits'. She said she picks companies where revenue grows roughly exponentially while headcount grows roughly linearly, as at Google and Stripe. She attributes part of the difference to AI giving many roles multiple-times productivity.
“there are a bunch of things where like, you know, the equivalent human is getting NX productivity.”
Using Tangelo, which is tied to the role assigned in Workday, all scheduling for a new hire's three to five weeks of onboarding is generated automatically. Weekly knowledge-check tests run through AI, and managers get automatic alerts on progress. The early alert system creates a forcing function for a serious conversation with the manager about whether a hire will make it.
“all of your scheduling for your first, whether it's three to five weeks in onboarding is done now straight up through AI.”
Jo Massie said the customer voice team of four could no longer be afforded, so she asked the team to deliver the same outcomes with about 1.5 full-time employees. She said the team did amazing work with AI and exceeded her expectations completely. She said the program's output feeds the product team's roadmaps.
“So earlier this year I was like, we can't afford to have 4 full time employees doing this anymore.”
Richmond said he agreed that copilot-style applications give one person much broader depth of knowledge. From an operational leverage standpoint, his aim is efficiency, meaning people can cover many more accounts because they have a clear picture of what they need to do and tools to do it very efficiently.
“I want, I want people to be able to to, to cover a lot more accounts because they've got a very clear picture of exactly what they need to do”
Wayland notes that friends from Salesforce are now at Qualified and that many companies are building AI SDRs. He says he no longer needs the same SDR capacity because AI tools can do some of the work, which allows a different SDR-to-AE ratio. He frames the choice as an open question: take the change as savings, or use it to fund initiatives Box previously couldn't afford.
“I don't have to have the same SDR capacity that I did before because I can use AI tools. So then if I can have a different ratio between SDRs and AEs, do I take that as savings or do I take that to fund other initiatives that maybe I couldn't afford before?”
said Amanda Kahlow (Topline), Gaurav Agarwal (Topline), Kyle Norton (The Revenue Leadership Podcast), Pablo Dominguez (Topline), Wade Foster (Topline)
Amanda Kahlow describes a superhuman as an orchestrated system rather than a face on a chatbot. One agent handles small talk, one runs the live demo, one brings up slides, and another builds a perspective slide during the call. She says the hard part is unifying these agents into one experience with rich context, and that a basic chatbot with a face is easy to build.
“each superhuman has roughly 15 agents working behind the scenes”
He described breaking a workflow into smaller processes, with each agent handling a narrow job, rather than building one agent to do everything. He presented this as the current practice at ClickUp, which he said is still learning as it goes.
“we try to limit an agent to three to five things that they should try to do, not more than that.”
She describes a demo model, a tool-calling agent that works out who the buyer is, and a conversation agent, all coordinating and passing the baton. She describes it as a duck: smooth on the surface, busy underneath.
“there's probably eight to 15 agents slash, you know, models that are working simultaneously to make the one superhuman talk.”
Kyle compared this to a prime number that cannot be divided further. Once an agent does that unit, he said, you have unlimited effort and unlimited supply of that job. He described going piece by piece through the revenue function at this level before scaling, and Jordan agreed that defining the most leveraged atomic task is key.
“you want to go down to like the smallest unit possible of that work”
He said a single agent doing everything would not be modular for others to use. Another Insight team could adapt the existing research agent for limited partner fundraising, or reuse the Gong agent for HR call checks, rather than starting from scratch.
“we highly recommend you build Modular pieces of a system versus having one massive agent just do everything”
He said there is no function where AI cannot be applied successfully, such as HR, finance or sales. Coding agents get a lot of discussion because they are quite good, but he said Zapier has evidence across basically every function and vertical.
“we have evidence in basically every function in Vertical where you can apply AI and be successful.”
Wade said tasks like this need to be broken into their smallest actionable components. Those components are then tied into a mix of deterministic and agentic workflows. He said this breakdown is what gives the advantages available today.
“these tasks are too big too abstract and is you do need to break these things down into their smallest actionable like components”
said Alex Bilmes (Revenue Builders), Aviv Canaani (The Revenue Leadership Podcast), Kyle Norton (The Revenue Leadership Podcast)
Alex Bilmes said customers are building what he calls a CRM hygiene machine. The service listens to all calls, looks across the data and updates the CRM so it stays accurate and can change every two minutes, meaning nobody has to update the CRM by hand.
“you can basically stand up an orchestration service it just goes listens to all the calls looks across all the data and updates your CRM”
Aviv says the AE used to choose the closed-lost reason, which was often inconsistent. DataRails now sends closed-lost customers a call, offering $50 to $100, in which an AI agent asks the questions and fills in the Salesforce fields. He says this gives the company the answer from the customer rather than the rep.
“Now we automate like calls going to our closed loss saying like okay we'll pay like $50 $100 just get a call you speak with AI”
Kyle Norton says his team first had reps approve or edit what Momentum synced into Salesforce. The team compared the information reps entered with what Momentum captured against transcripts, and found Momentum more complete and more accurate. They then let Momentum write all the fields, and reps can edit them if they see an error.
“we just flipped it to like no, like Momentum's just going to write all the fields. We're never going to ask the reps.”
said Kyle Norton (The Revenue Leadership Podcast), Mike Carpenter (Topline), Justin Shriber (Topline), Alex Bilmes (Revenue Builders)
Norton says the model fills gaps and makes things up because it is optimizing for an answer the user will thank it for, and the user often doesn't know enough to give accurate feedback. He presents a structured harness or harness platform as the fix for getting correct answers.
“if you just, you know, MCP your Claude code instance or your Claude desktop instance into a bunch of things and ask it questions, it's, it's lies. It lies to you. It just, it fills in the gaps.”
He said leaders increasingly present Claude analyses that data science teams then find wrong. He named Vasco as the most interesting tool he has seen for getting correct answers, citing the effort needed to get the right context into the model and stop it answering when it doesn't know. He built a CRO daily in it.
“that's a space that that is going to be more and more of a topic of conversation pretty soon, which is like data harness.”
Mike said systems like Claude generate new code every time they run a query without pulling results back into memory to learn from them. He said a slightly different wording can produce a different answer, delivered with certainty. He presented this as a reason the output cannot be trusted without checks.
“It's not pulling anything back into memory and learning from it”
Mike said a general LLM will index Slack, email and documents and learn wrong data. He said the approach needs a co-pilot or digital twin that maps potential links, with a large human intervention piece. He said Xfactor has staffed up its intelligence work so the system can learn those cause-and-effect relationships.
“You need causal AI, and you need a co -pilot, a twin that starts mapping potential links.”
Mike said an AI that gives answers with certainty will lead companies to pivot their business in ways they do not know are wrong. He said they will not find out until numbers start dropping. He described most of the calls his company gets as companies that added a cloud model to their data without a data lake or tests.
“you're going to pivot your business in a way that you're not going to know what's wrong until your numbers start dropping”
Mike said his team ran Claude on the same dataset they use and it completely hallucinated an answer. He said it did not simply drift but delivered the answer with certainty. He called this confident wrong answer his biggest competitor right now, pointing to companies that put an LLM on top of their CRM.
“we've taken Claude and just thrown it on the same dataset we have, and it completely hallucinates an answer, but it doesn't just hallucinate or drift an answer. It gives it to you with certainty.”
He said that although anyone can now build MCPs into any system, this approach gives the LLM one piece of a puzzle and then asks a complicated question. He said the quality of the answer is therefore uncertain, which is why he described Terret's architecture as a revenue graph rather than a set of connectors.
“Today, you can build MCPs in any system that you want, but that's not the answer”
On security, he said AI should not be able to backdoor into sensitive systems such as CRM or a data warehouse and expose data inappropriately. On accuracy, he said that asking for a win rate produced a different number each time because the LLM connected systems inconsistently, and on scale he said inefficient prompting burns tokens faster than the value produced.
“So it was being able to build a revenue graph that one was secure, two was accurate, and three that was scalable.”
Alex Bilmes said that systems like Salesforce, Zendesk, enablement platforms and Gong were not built for agents, so a layer that turns company data into knowledge agents can act on is likely needed. He said teams are turning off traditional enablement systems frequently once methodology and content can be connected to an LLM.
“We're seeing teams turn off of traditional enablement systems pretty frequently.”
Alex Bilmes said an LLM context window is very small, and most companies connect Salesforce, Gong, Slack and Google Drive directly to a model without first building a foundation. His company ingests every call, identifies objections, pain points and domain-specific facts, and condenses them before the model queries them.
“you end up just throwing a ton of a ton of noise at an LLM and it doesn't really know what to do with it.”
Schuck said CRM data is necessary but not sufficient for go-to-market AI. He said the data layer above the CRM should bring in unstructured sources such as conversations, emails, calls, earnings calls, conference transcripts, job postings and press releases to identify key moments and signals. He said without this layer you cannot deliver go-to-market AI or the next best action for a seller.
“there has to be a data layer above the CRM that masters the CRM data and then brings in all of that unstructured data”
said Simon Farthing ([Un]Churned), Adrian Rosenkranz (The Revenue Leadership Podcast), Kyle Norton (Topline), Jordan Crawford (The Revenue Leadership Podcast)
Simon said that if a tool is used by an individual, building it yourself is awesome, but if more than a handful of people are using it, you really have to go with the vendor. For products sold to a customer base, traditional software engineering methodologies remain as important as ever.
“If it's an individual tool, building it yourself is awesome. If more than a handful of people are using it, you really have to go the vendor player.”
Simon said in-house building, like Mission Control, needs no engineering resources or product backlog and lets you shape and iterate quickly, so speed is its advantage. For the middle layer he would tap into vendors such as Gainsight, Vaughn or Claude, where enterprise connectors, memory and infrastructure matter and an individual vibe coder cannot rival what is bought. Products sold to customers, such as Bloomreach's Lumi, he said should be built and owned by engineers under traditional software methods.
“it requires no engineering resources no product backlog”
He said governance, roles and permissions should be bought so teams can build AI on top of them rather than rebuild and maintain them repeatedly. One-time analyses can be lightweight artifacts, but recurring needs should become applications with a proper view. He also asks why building beats buying, weighing it against price relative to value.
“You need to buy systems that you can build with AI on top of it.”
Kyle Norton says he has not seen sales engagement providers build an AI product that does real intelligence work such as next-best-action or deal recommendations, and asks whether the hosts have seen otherwise. He says to output internally built signals such as a deal health score or estimated win rate onto the lead record, using the engagement platforms as the place to display them. He says he would not vibe-code the daily rep tool because it needs rock-solid stability and a support ecosystem.
“I haven't seen any of the sales engagement providers build an AI product that does any of the intelligence stuff.”
He said Claude Code was a better tool for building and testing the original prompts. He said Clay is the better place to run them for ongoing use because it is editable, pushes to the user's systems, and gives observability through its interface.
“the best place for you to deploy them is in clay”
said Asad Zaman (Topline), Mark Roberge (Topline), AJ Bruno (Topline), Joubin Mirzadegan (Grit), Steve Cox (Topline), Jeremey Donovan (The Revenue Leadership Podcast)
Asad Zaman said he works with many of these companies and that they are still figuring out how to use AI in their departments. He said engineering can look like magic, but going to HR or go-to-market shows that it is not that different from other companies. He cautioned against treating AE productivity as proof of internal AI skill, since it may reflect demand and high close rates from experimenting customers.
“I can tell you that they're not as AI pilled internally as you would like to believe.”
He said that when operators such as Kyle Norton put out their AI work and he calls them to see it, even they admit some of what they describe is still aspirational. He said he thinks more evidence should appear by the end of this year.
“It's a disappointment of across our entire portfolio and it's not I think it's not our portfolio It's just the entire ecosystem.”
He presented the figures from ICONIQ's state of go-to-market study released in January 2026. He noted that the study used 2025 data, and that AI tooling has changed quickly since then.
“AI clearly lifts top of funnel, 11 % on the lead to MQL and MQL to SQL another 8%, but barely moves active deal cycles.”
Responding to Ryan's question about what he is seeing, he named engineering as the first function where AI tools have raised output sharply, and said he has also seen it in legal and support. He said he thinks sales is about to see the same kind of impact on productivity, but that it has not happened yet.
“Like in engineering, we are getting people from 15 to like 50 points a game with these tools, maybe more.”
He uses this study to explain why many companies he talks to have moved from pilots and tests to questioning value. He argues that the current stage is a return to conversations with incumbent software vendors about what they are doing with AI.
“there was a recent study done by MIT that said, 95% of AI trials have ended in not being able to show proven ROI and actually not be cost efficient.”
Jeremey Donovan compared CROs at top-performing companies with CROs at average to below-average companies and found no obvious silver bullet in AI use cases. He described the findings as disappointing but said they reflect the current state of adoption from the survey and his conversations with CROs.
“there's not actually a huge difference in which use cases they're adopting and in which order.”
said Tomasz Tunguz (Topline), Rick Smolen (Topline), Marcy Stoudt (Revenue Builders), Sam Jacobs (Topline), Arvind Jain (Grit)
Tunguz said his total edit count held at 134 per post regardless of how long he had been working this way, while a graded review of ten years of posts showed a 20% quality increase in 2026. He still spends an hour to an hour and a half per post, but said the research, citations and depth of analysis improved. He compared this to chess grandmasters who keep training as much while holding themselves to a higher standard.
“if you're a chess grandmaster and you want to be better with AI, you don't train less. You train just as much, but you hold yourself to a higher standard.”
Rick says it is fine to draft an email with AI and then ask it to add more empathy, since he describes himself as naturally a bit of a jerk. He stresses that the message is still from the person who sends it and must be defensible. He sees AI as useful for copy in a thought partnership role, not as a substitute for the person's own work.
“I think that these things can be a thought partner, but you have to own your content.”
Marcy says she set custom instructions for her AI strategist, uploaded the strategy books she follows and her core thinking, and told it never to accept an answer based on what she says. She says the setup keeps her from feeling stale and pushes her to think bigger before calls.
“I uploaded the strategy books that I follow. I uploaded like some main core thinkings. And then I say have it challenge me never accept an answer based on what I say always thinking”
Sam Jacobs says he comes up with a thesis, uses AI for deep research so his ideas have credibility, and checks that the supporting research reports and papers exist. He says he does not read every source in detail but reads enough to see whether they support what he believes, and he calls this the best content he has produced.
“I come up with a thesis. I use AI to do deep research so that my ideas have credibility.”
He describes giving AI rough thoughts and asking it to structure them, which helps him get past writer's block or lack of time. He then works on the resulting artifact and adds his own thinking. He warns against work slop, where people generate large amounts of AI text and push it onto others to read.
“I will sort of give some rough unorganized thoughts to AI and say that, hey, put that in some sort of a structure for me and I get some artifact and then I start to work on it”
said Adam Liska (Topline), Kyle Norton (The Revenue Leadership Podcast), John McMahon (Revenue Builders)
Adam says that within engineering, much tribal knowledge is now being documented in the code base itself, so that when engineers and AI agents work on the code, the context is available to everyone. He says a similar approach should happen in go-to-market. Host Sam Jacobs adds that LLMs now let teams give an agent unstructured tribal knowledge as context, where before only structured data could be used.
“a lot of these tribal knowledge that was within the engineering team is now being documented actually within the code base itself. So that when engineers work, but also when AI agents work on that code base, all that context is available to everyone.”
Norton says that if you want to know where go-to-market is going, look at where product development is today, with its harness, context engineering and loops, and expects go-to-market to be there in about 18 months. Rutten separately expects software vendors to offer similar capabilities eventually, with a window of six to twelve months, perhaps eighteen, where building in-house is an advantage.
“If you want to know where go to market is going, you just have to look at where product development is today. And that's where go to market's going to be in 18 months.”
McMahon says software developers took their domain expertise and embedded it in AI tools, and that the same is expected to happen across other functions. He predicts that sales-specific tools built on domain expertise will emerge, though he notes that current tools are mostly general large language models.
“The reason I think that software development occurred first is because software development people know software development so they took their domain expertise and they took the time to embed it in an AI tool.”
said Ghazi Masood (The Revenue Leadership Podcast), Alex Varel (Revenue Builders), Christina Meng ([Un]Churned)
Ghazi said the app is the company's version of Gainsight, giving customer advocates and field engineers visibility into utilization, health score and engagement for each customer. He said the field engineer who built it took less than a day and that they keep changing it as the company grows.
“I think it took him less than a day.”
Alex describes Jimmy Lee, his RevOps lead, who trained five or six agents over Easter weekend. One agent moves customer feedback from JIRA tickets into product feedback, another supports forecasting, and another monitors solution architecture work, with six more in progress. Alex says this kind of setup should multiply productivity.
“He's got one for voice of the customers to product feedback going through JIRA tickets and things like that.”
Christina Meng said she is not a technical person and cannot code. She said she used Codex to build a web app on top of her hard-to-read spreadsheet of internal and external champion events, pulling the right information for internal teams and linking to event recaps in Slack, notes and landing pages.
“I was able to use Codex to help me build a web app that essentially pulls the right level of information for our internal teams”
said Simon Farthing ([Un]Churned), Stuart Gwynn (Revenue Builders), Adnan Rahman ([Un]Churned), Mark Roberge (The Science of Scaling), Andy Shorkey (The Science of Scaling)
Preparing for an exec meeting used to mean switching between about six open tabs, talking to the team and testing insights, which Simon estimated took three or four hours. With the tool he gets the same prep in about 10 minutes, which leaves the rest of the time for interrogating the insight and building the talk track.
“I now get that in 10 minutes, and then I can spend the rest of the time really interrogating the insight”
Stuart says he uses AI to research accounts, grab information from internal systems, and draft memos. He gives the example of a memo tied to a prep call with MongoDB's CEO, which used to take hours to write and now takes minutes. He says he does not run with the AI output and instead augments it. He calls this table stakes and says he wants to use AI as a force multiplier to keep scaling his own work.
“That used to take me hours to write that memo and now it's taking me minutes”
Adnan says Copilot provides call prep insights and helps CSMs prepare for calls with less manual work. He also says AI helps formulate messages in a customer-friendly way. He describes this as an efficiency gain the team is already getting.
“You know, we use co-pilot within GainSight. That's great for call prep, insights, preparing CSMs to cut out some of the manual administrative work of preparing for a call.”
Mark suggests asking an AI the same open-ended questions you would ask in the meeting: about the person, their company, and their organizational and personal needs. He also suggests asking how they would evaluate your company and your competitor, how to sell your product against the competitor, and how the competitor would sell against you. He says this makes reps far more prepared than was possible before AI.
“I think just coming back to like, what are you going to do in the meeting and just do it now?”
Andy Shorkey says a rep tells the app which account they are meeting, using the example of a sporting goods company. The app then lays out a point of view on the customer, the customer's orientation toward AI, and five areas mapped to Writer's capabilities. It produces a solution map, the point of view, and a script or question set tailored to each executive's role. He says it essentially gives the rep the first call.
“And so it essentially would create a solution map. It would create the point of view. It would also create the script or the questions for the executives, depending on their role and responsibility.”
said Simon Farthing ([Un]Churned), Parag Agarwal (Grit), Jason Goldsmith ([Un]Churned), Margo Martin ([Un]Churned)
The tool, which Simon calls Mission Control, connects customer health, one-on-one project tracking and his other priorities, and plugs into Bloomreach's core tools. He uses it to prepare for exec calls and to understand customer health, and says it helps with the repeatable parts of his job faster and smarter because the context layer is improved.
“So it's a personal command center that I built in Claude.”
He built it partly as an experiment to see how much a machine working on his behalf spends on Parallel's web access versus internal data. It knows his written top three priorities, acts proactively and is 'paranoid on my behalf,' which makes him more deliberate about which balls he drops. The host said Agarwal had spent $378 on tokens by 1 p.m. one day. Agarwal calls the original version 'outrageously inefficient'; after optimising it he expects to spend about $100 to a couple hundred dollars a day.
“it makes me feel more secure that I am not dropping the important balls because it is also paranoid on my behalf. It is inferring my priorities.”
He is invited to more meetings than he can attend or review. The agent looks at the previous day's meetings and tells him what he missed that needed attention, which he calls a yesterday recap. He uses it to plan the day and the rest of the week and to filter out background noise.
“I've got one that just goes and looks at yesterday's meetings and says, okay, this is what you needed to pay attention to that you didn't hear directly”
She said senior leaders get caught in the tyranny of the day, and her agent keeps her aware of meetings, prepared for the week, and reminded of follow-ups with customers. She gave an example of it reminding her of a podcast recording she had forgotten. She said its most important value is keeping her on task.
“my little chief of staff keeps me on track every day”
said Jason Goldsmith ([Un]Churned), Kyle Norton (The Revenue Leadership Podcast), Josh Schachter ([Un]Churned)
He encouraged his team to explore, experiment and make mistakes with Claude and other AI tools in their day-to-day work. A dashboard tracked Copilot engagement across the team, and his name was at the top. He said the measure might not be perfect, but it showed his team that Copilot could be used beyond summarizing recorded meetings.
“we had like a dashboard that would get sent out for this was on first using co -pilot”
The speaker says that if you are a VP, CMO or CRO and have not tried to create an agent or wrestle with these tools, the world will pass you by. He says this is not a moment to re-delegate AI to your RevOps team and sit back. He says you need to be hands on keyboard a little.
“if you are a VP or a CMO or CRO and you have not tried to create an agent or truly like wrestle with these tools the world will pass you by”
Jeremey Donovan described a CRO he interviewed who was personally writing many prompts. He noted some might argue that is not the best use of a larger-company CRO's time, but said the reinforcement is part of the discipline. Kyle Norton said he would argue it is a great use of time and that he is moving to a Claude Enterprise plan mainly for centralized skill management. That way he can update skills in one place instead of copy-pasting markdown files, and everyone using a skill gets the improvement.
“my most important feature is centralized skill management.”
Josh says Leeron has built a set of agents, which she calls her little people, and tinkers with them on her Fridays. Josh says this has let Leeron scale herself and show her team. Josh says they spoke to Leeron months earlier.
“she's just got like all of her little agents that that she just tinkers with on her Fridays”
He says these leaders vibe coded on weekends and nights and compared agent spaces so they could show their teams rather than only tell them. He says adoption can also start bottom-up from passionate ops people and CSMs, but top-down leaders need to carve out that time for themselves.
“These leaders, they were leading by example in like diving into AI themselves.”
said Alex Mashrabov (The Twenty Minute VC), Joubin Mirzadegan (Grit), Parag Agarwal (Grit)
Mashrabov says creative staff began vibe coding. One person spent over $30K in a week building an asset-organization workflow that wasn't production-ready but taught the team a lot, and many spend over $10K a week. He acknowledges his finance team thinks he is too stubborn about not controlling the spend, but calls the experience net positive. He expects top '10x' engineers and creatives to reach $50K-100K a month and to ask for comparable salary raises, while spend for functions like legal and finance stabilizes quickly.
“I do believe we are gonna get to spend close to 50k and 100k a month for those who can call 10x engineers, 10x creatives.”
Joubin Mirzadegan says his company Roadrunner ran 'let it rip' for about six months, then moved all spend onto one credit card to see where it went, and then tracked which models it went to. They track spend but enforce nothing. The one rule of thumb is that an engineer spending their salary on tokens 'better be pretty good.' He would rather pay more now so his best engineers can stay on the bleeding edge, for example chaining five agents together.
“the short answer is we track it, but we don't enforce anything today.”
Because no one knows the right level of token spend at any given time, he asks whether he would rather be wrong by being too conservative or by spending too much time and tokens with models. His personal choice is to err toward token maxing. He says you will never be exactly right, but knowing which way you prefer to be wrong tells you what to do.
“if you know which way you'd rather be wrong, that tells you what you should do.”
said Jonathan Moss (The Revenue Leadership Podcast), Tim Rutten (The Revenue Leadership Podcast), Kyle Norton (Topline)
His two objections were lock-in, which he sees as a big risk while things are moving fast, and giving a vendor the company's business logic, processes, context and IP. He says some parts of the stack should be bought, but the brain and harness are where to spend time building. Norton noted that the in-house technical capability to do this is not the norm in his experience.
“we were about to put all of this into a third party. And I was, and, and I was like, whoa, whoa, whoa, before we do that, how Are you going to port it over to something else now?”
Backbase did not renew 6sense because it did not fit its motion. Since 6sense informed campaigning and parts of the engine, they had to rebuild within two months. Rutten says the team built a white-box signal engine with a scoring rubric they tune themselves, which he says gives them go-to-market alpha once it connects to other building blocks.
“we completely built a signal engine bespoke that is not blackbox but white box I know exactly what is happening there”
Kyle Norton says the core intelligence of the organization needs to be built and managed centrally, and he doesn't think it should be bought. He says the things to buy are workflows and user experiences. He says build-versus-buy was a big topic of conversation at the Clay CRO summit he attended.
“I think you have to own the intelligence. You have to build the intelligence internally and buy things that are like workflows and user experience.”
said Jeanne DeWitt Grosser (Grit), Aman Narang (Grit), Diego Ballona ([Un]Churned)
Grosser said Vercel gets a lot of support volume and 91% of it is resolved by an agent. The remaining support team, which is technical, now focuses on legitimate product problems, works closely with engineering, and in some cases writes pull requests to fix issues itself.
“Vercel gets a lot of support volume, as you'd expect, and 91% of it, we resolve with an agent.”
Aman said Toast has a product called ToastIQ, built on Toast's data and accessible through a chat interface that can answer questions and make changes to the back end. He said the support agent is partly off-the-shelf software from a partner, and Toast is still using foundation models in parts of its AI work.
“We're doing about half our support through it now, or through the support agent.”
Diego said Intercom has nearly 7,000 customers using Fin, and that the average outcome across all segments is a 67 percent resolution rate.
“the average outcome that our customers get it is 67% resolution rate.”
said Mark Stiving (Impact Pricing), Mike Carpenter (Topline), Pablo Dominguez (Topline)
Mark Stiving says he liked that the AI gave the entire justification for the price, so he does not have to search for the relevant pieces himself. He imagines overriding a price for a personal reason, such as giving a brother-in-law a deal, and says the AI could then ask why and learn from the answer.
“It actually gave me the entire justification for why it created that price.”
Mike said the first question from clients is usually that the results do not match their spreadsheets. He said Xfactor now puts its methodology in front of customers so they can manually check how the results are calculated. He said building consumer confidence was the biggest challenge.
“But now we put it in the methodology so they can actually go manually look at the methodology that we use to calculate those results.”
Mike said Xfactor runs many checks on each finding, and if it cannot back a result with certainty it drops the rating to 60 percent. He said that rating tells the user to look at the finding carefully and test it slowly. He said removing statistical error from the system was one of the company's biggest issues.
“if we can't back it up with running tons of different sum checks about it with certainty, we'll drop it to a 60 % probability rating”
He said that without this instruction, an agent might pull a market share figure from articles and present it with no source. Pablo said that across the first ten diligences he has seen very little if any hallucination.
“You do not provide a number. In fact, you say I do not have an answer”
said Saumyo Mukherjee ([Un]Churned), Kellie Snyder ([Un]Churned), Chael Banks ([Un]Churned), Jeremey Donovan (The Revenue Leadership Podcast)
Saumyo says that with AI tools in everyone's hands, adoption is much more democratized, which makes consolidation harder. Braze has rolled out tools such as Claude and Gemini to staff, who are free to vibe code. He describes the chaos as good because creative ideas flow and the team learns from what people build, but consolidation and guardrails have to follow.
“the good part of it is creative juices flow And we kind of get to learn a lot of cool things that people are doing, and we get a lot ideas that we get to pick from and then consolidate.”
Kellie said that people given AI tools will vary, with the best people doing the most interesting things and others feeling scared. Only recently has everyone on her team been given Claude access, and the team is starting to share lessons. Her aim is that the agents' work becomes a standard everyone knows and uses day to day.
“the commonality for me is capturing those wins and creating consistency out of it, right?”
Chael says the problem with building too many tools is that the result has to be scaled consistently so the experience is the same across all customers. He says that scaling in this way lets the company keep best practices embedded in the process. He describes this as the reason Okta is consolidating its many prototypes.
“The problem with building too much is you have to scale that through in a consistent way.”
Chael says there is a supernova of AI building across the world and that his team has many ways to solve the same problem. He says he is comfortable with that level of innovation because the best approaches will surface and can then be brought together. He says the broader challenge is that everybody has the power to build and the responsibility has not yet been worked out.
“we have nine ways to solve the same problem.”
Jeremey Donovan said that in his conversations, the more successful CROs let reps and others test AI ideas from the bottom up, picked the best ones, and then built a mechanism to syndicate them across the team. He said bottom-up tinkering alone only goes so far, which is why he expects more top-down AI work.
“they've developed a system to farm ideas from the bottom up”
said Jeanne DeWitt Grosser (Grit), Christina Parra ([Un]Churned), Mark Vovsi ([Un]Churned), Manish Chawla ([Un]Churned)
Grosser said Vercel has a data science agent you can ask anything you would have sent to RevOps. As a result, she described the RevOps team as leaner and more efficient than teams she had in prior roles.
“We have a great data science agent that, like, I mean, you can ask it anything you would have sent to RevOps.”
Christina Parra described Data Scout as like having a data analyst sitting next to you who answers natural-language questions. As a CS leader she uses it to uncover insights about her business quickly. When she does need a data scientist, she arrives having already tried and run the data, which she says builds trust. She said multiplayer deep-agent use cases like this are where Notion will keep focusing, internally and with customers.
“imagine there's a data analyst sitting next to you.”
He described a case at end of quarter, when his CCO asks about a customer before a meeting with the CEO, a customer call, or a large renewal. Mark said the CCO does not log into the CSP and wants the answer very quickly. He presented this as one use case for accessing CS data headlessly.
“They want to access that information very quickly, right, be it on Claude or a tool like that.”
He said he has become more disciplined and now asks Slack bot for data rather than asking his team to pull it. He said this does not transform the function immediately, but he sees it as a way to create more productivity.
“So I've also become a bit more disciplined rather than ask my team for I have a question. I'd like to see the data for XY and Z, I asked Slack bot.”
said Ross Simmonds (The Dave Gerhardt Show), Eric Gilpin ([Un]Churned)
Ross calls Reddit 'ridiculously valuable' still. He says people don't realize how much B2B dealmaking it shapes, because it is such a powerful source of training data for AI tools like Gemini. Manipulating and spamming Reddit is the 2026 marketing trend he hopes ends.
“There's billions of dollars worth of deals in B2B that are being influenced by Reddit that people don't even realize because Reddit is such a powerful training data, platform.”
Ross says Foundation created the piece for a client about three years ago. Instead of publishing it once, they kept updating, distributing and promoting it. It now appears in LLM answers to some of the highest-priority prompts in the client's category, and the client still generates leads from it.
“We continue to update it, distribute it and promote it is now cited in the LLM consistently across some of the most priority prompts in their category, and they're generating leads on the back of this asset from, like, three years ago.”
Ross says LLMs pull from Reddit, SEO and the other channels that organic distribution targets when they answer questions. He says demand for Foundation Marketing's services has increased as a result. He also launched Distribution.ai as software for marketers and sales executives to distribute their stories at scale.
“Ai is now using all of those different distribution surfaces to translate back into the LLM answers to questions. So our services have increased in demand because of that.”
Eric said tracking the prompt queries behind citations helps the AEO team learn and test. He said G2 has a big push for listicles and best-of mentions because they are ranking high in AEO results right now. He said they would not know this without tracking their own AEO performance.
“we have a big push for listicles and like best of mentions”
Eric said that if 55% of buyers are starting in LLMs, a B2B software company that does not let LLMs scrape its site is invisible. He said LLMs currently love reviews because they are conversational, vetted and moderated, and G2 does moderation and verifies reviews itself.
“if 55% of all buyers, which is up from 25% of buyers the previous year, are starting there”
Eric said that after GPT-3 was released in November 2022, G2 had to decide whether to make its public data available to LLMs. He joined in September 2023 after the decision, which previous leadership made on the belief that the future always wins and that dollars follow eyeballs. G2 now lets models train on its data daily and ships data in the formats they want, which he said drives citation traffic and brand influence.
“And so did we allow LLMS to scrape all of our public data or not?”
said Asad Zaman (Topline), Christopher O'Donnell (The Science of Scaling)
He cited an investor who builds heavily with AI saying his agents love working with Clay. Asad reasons that if three companies do the same thing and agents prefer one, you have to be bullish on that one, at least for now. He also pointed to Clay's long history and pivots as giving it time to build strong operations.
“if that's the case, at least for now, you have to be bullish on the company that the agents love working with.”
Christopher argues that humans no longer judge a tool by its interface or feature set, because the AI is the one using it. He suggests connecting a set of tools to Claude, asking what it could use those MCP servers for, and having it try them and report back. Some excellent products 'won't really give Claude what Claude really wants.'
“You can connect a bunch of different tools to something like Claude and say, what do you think of all of these? Like, what could I use these MCP servers for? Try some of it.”
said Jack Altman (The Twenty Minute VC), Jake Saper ([Un]Churned), Tim Rutten (Topline)
He said there is real anxiety about non-American models. Even so, he expects enterprises to build on open weights, use inference providers to make their own models, and then run those models themselves. He said he does not know how much impact this will have.
“I also think we are starting to see and will continue to see a lot of enterprises post train their own models. and draft off of open weights and use the inference, you know, companies to make their own models and then run them themselves.”
He says open-weight models can be hosted privately so company data stays private, and that they are a big boon for AINS businesses delivering services more cheaply. He mentions a letter Emergence co-authored with NVIDIA and Microsoft, which his partner Gordon helped write, supporting the rise of open-weight models.
“open-weight models can make that token spend exponentially cheaper than it was even a few months ago”
He said this is logically what banks will do because they want to contain risk and control unit economic cost, and a few steps are already underway. He said the bank should stay model neutral and choose different models for different workloads. He framed the on-premises move as eventual rather than current.
“They want to contain risk. They want to control the unit economic cost.”
said Jason Lemkin (The Twenty Minute VC), Alex Mashrabov (The Twenty Minute VC), Anastasios Angelopoulos (Grit)
Jason Lemkin said any eval can be made to look great. What matters to him is speed, real outputs, and performance on real and mission-critical workflows, such as correctly stating whether an item is in stock. He said evals focus on things like 'chess championships', so he would want to try Beam himself, for example on OpenRouter, before believing the claims.
“I found that every single eval needs like three asterisks and four daggers next to”
He says researchers at larger labs have told him that test data gets put into training and other tricks are used to hit benchmarks for quarterly bonuses. He attributes this to incentives at big companies. As evidence he cites OpenRouter data showing Google as the only relevant US incumbent, while in China, where he says benchmark obsession is probably lower, Tencent, Xiaomi and Alibaba are relevant. He admits he himself once wrongly chased benchmarks.
“what happens is that they start to put test data into the training.”
Early on, the models that scored well on multiple-choice tests like MMLU were not the ones people liked to use, because 'they all train at the test.' He cites the viral 'pelican riding a bicycle' SVG benchmark, which he says became saturated once it entered training data. His view is that benchmarks are invented because they attract attention, and that they say little about how a model does on an actual workflow.
“the model that does well on the test is not the same model that I like to use. And that's doing well in the real world. It's because they all train at the test. And so we had this philosophy that it's really about reality.”
said Jen Igartua (Topline), Christopher O'Donnell (The Science of Scaling), Rick Smolen (Topline)
She describes the matrix as skills and taste versus working with or without AI. Highly skilled people using AI do a really great job, and low-skilled people using AI produce slop. She says this gap is showing up much more now.
“if you have, you're very skilled with AI, you do really great job. If you're low skilled with AI, you're a slop cannon.”
He describes two groups: people who are 2-5x or more productive, and a large group still at baseline. The 15% figure would describe someone using autocomplete while still coding by hand, which he says is not what anyone is really doing.
“My take on all that noise is it's going to be pretty bimodal already.”
Rick says the spread between top and bottom performers will widen rather than lift everyone equally. He says the best sellers will use AI to prep, target better customers and follow up more efficiently, while those who expect AI to do the work for them will be unmistakable. This is his prediction, based on what he says he is observing.
“I do believe that the best will become 10x better.”
Rick compares AI to debt, which amplifies results in either direction. He says AI is not good or bad in itself; what individuals and teams do with it determines the outcome. He describes using AI as a partner that helps a seller prepare, target and follow up, while the seller still owns the work.
“It's like debt. Like debt is leverage. It can make performance exceptionally good or it can bankrupt you.”
said Gaurav Agarwal (Topline), Cassie Vaughn ([Un]Churned), Brett Queener ([Un]Churned)
Gaurav said his mindset moved from using AI as a sidekick to keep existing structures to AI doing the work. He argued that humans will build AI to do the job, and that AI will do it better than an 80th percentile human. He described the resulting human role as managing and training AI.
“AI will do the job better than an 80th percentile human. And then our jobs become managers and trainers of AI.”
She said agents can be trained on repeatable playbooks and frameworks, and in some cases will deploy themselves based on customer interactions. CSMs would then focus on judgement calls and relationship-building. She said agent management is a skill that has to be trained, since agents do not simply run on their own once deployed, and said she believes CSMs are uniquely equipped to take this on and that monday.com is currently doing so.
“shifting the role of the CSM to think about being an agent manager.”
He describes agents as semi-autonomous employees: rather than giving step-by-step instructions, you assign an outcome, and the agent builds a team to do the work. He says the human's role shifts to review and management. He presents this as the new era he believes is underway.
“you assigned an outcome. It can be a big outcome. And then you tell it to build the team to work with you, to go do the work.”
said Marcy Stoudt (Revenue Builders), Kristi Faltorusso ([Un]Churned), Mark Roberge (The Science of Scaling)
Marcy says she commits to 20 minutes a day of AI learning and subscribes to a number of newsletters, many of which are vendors selling to her. She used to call her mentor Edwin about such offers. Now, when something is interesting, she copies it into her main LLM and asks it to teach her how to do it.
“Copy paste uploaded into your main go -to one and say teach me how to do this”
She says she used YouTube and TikTok, where 5 to 10 minute clips were ideal, writing notes, trying something, then going about her day. She says she also spent time building. She encouraged her team, and some members embraced it more than others, and she expects those who did to become breakout stars.
“I was allocating an hour a day to learning”
Mark says people in sales, marketing, HR or finance should get down to the foundational tech level by playing with GPTs and AI apps at the configuration level. He suggests carving out an additional 25% of time to try automating something you do. The first few reps will take longer, but if you can cut the time for that task in half over the long term, he says that is huge. He cited Writer CEO May Habib's ability to go deep on LLMs and the stack as a selling founder as an example of this technical depth.
“They carve out an additional 25% of time to see if you can automate something that you do. It's going to take more time, the first few reps.”
said Manny Medina (Topline), Amanda Kahlow (Topline, Revenue Builders)
A host argued that even frontier models fail on 3% to 5% of multi-step tasks, which limits what enterprises can trust AI to do, even with forward-deployed engineers on top. Manny replied that teams had the same problem before and that AI isn't a magic bullet. Frequent misfires usually point to a harness, memory or context problem that you debug until it's fixed. He says the goal is an organization that needs fewer people, has more throughput or delivers higher quality.
“when you're seeing a lot of errors or a lot of misfires, then you may have a harness problem, maybe you have a memory problem, maybe you have a context problem”
Amanda Kahlow says the AI does not need to be perfect because human sellers are not perfectly accurate either. She contrasts go-to-market with medical research or financial services, where a 5% error rate would not be acceptable. She says this is why the approach works well in sales.
“I just have to be as accurate as a human. The bar is just low basically.”
She says sellers hallucinate often on calls, and that no customer has reported a superhuman hallucinating in a way that hurt their business. She says answers cite where they came from and that the agent is held to tight guardrails.
“It's not that they have bad intentions. They're just trying to move the deal forward”
She asks how often sales reps hallucinate and says they do it nefariously, knowing they are doing it to move deals forward. By contrast, she says, as soon as AI hallucinates once, you can be sure it is never going to make that mistake again.
“I mean, do your sales reps hallucinate? And how often do they do it?”
She said she asks people whether their sales reps hallucinate and how often they do, and she claimed reps sometimes do it knowingly to get a deal done. She said that when AI hallucinates, its maker must make sure it never repeats the mistake. She said trust in AI is still crossing a chasm.
“do your sales reps hallucinate? And how often do they do it?”
said Saumyo Mukherjee ([Un]Churned), Sam Jacobs (Topline)
Braze's quarterly hackathon, run by a peer, is open to GTM teams (which include success, support and services) as well as non-GTM teams, and aims to include new people each quarter. Day one collects use cases, defines the problem and pressure-tests how business critical it is. On day two the solutions are built, with engineers helping requesters. Outputs are proof-of-concepts, awards are given, and the results are reviewed in the weekly AI meeting. The top two, a propensity to renew score and one other, are planned for productionizing in Q2.
“Out of that, it's a two-day hackathon where we kind of day one is collecting use cases and defining the use case as defining the problem.”
Sam Jacobs suggests leaders might give people a framework with timed steps: the first 30 minutes listing activities and actions in the job, the next 30 minutes mapping those to things a computer or agent could handle, and then seeing how far they get trying to build it. He says if people get to that point, it is probably a lot more powerful and immediate for them. He offers it as a suggestion for leaders to give their teams.
“maybe give them a framework like let's you know the first 30 minutes you're gonna list activities and actions that you take in the course of your job and then the next 30 minutes you're gonna map those to things that you think a computer or an agent could possibly handle for you”
Sam Jacobs describes Tim Rath running a 24-hour hackathon in which his team was broken into four squads that then talked about what they built. Sam says that once people realise much of the work is just speaking English to a computer, it becomes far less intimidating, and that the things you can build out of the box with these tools are quite incredible.
“So he broke people up into four squads, they had a hackathon, they talked about what they built.”
said John Gilbo (Impact Pricing), Jason Goldsmith ([Un]Churned)
He says incoming client data is always an issue and a common slow point. Agentic tools sit on top of the implementation process to streamline it and find data errors. He gives this as one reason implementations are faster and less expensive than in the past.
“one of the reasons our implementations are faster and not as expensive as in the past. We have a lot of agentic tools sitting on top of that process to streamline it, to find errors and data is always an issue”
Jason Goldsmith said his technical implementation staff used to take six to 12 weeks per cycle to build an integration when coding by hand. When Deltek began experimenting with Claude Code, those timelines fell by 50-60% depending on complexity, often by half or more. He called custom work the long pole in an implementation and said the speedup let the services organization deliver to customers faster.
“it would take six, eight, 10, 12 weeks to go through a cycle to develop, uh, you know, an integration, right?”
said Sam Costello (Revenue Builders), Jeremey Donovan (The Revenue Leadership Podcast)
Sam Costello says buying processes now include AI committees that review all investments, and AI addenda that may be needed on existing contracts. As an illustration, he describes how a deal forecast in commit, with an MSA already in place, could slip if purchasing finds no AI addendum with two weeks left in the quarter and sends it into a legal process. He says the team triangulates with peers to find such pitfalls, though it is still difficult.
“there's now an AI committee that wants to review all the investments”
He said the companies he supports range from zero to 100 million ARR, while other Insight teams cover companies above 100 million ARR. He described two dimensions, company size and whether a company serves regulated industries, and said security and privacy need to be included as partners. He separates AI inside the product from AI for operational efficiency and growth when discussing governance.
“absolutely as the companies get bigger and certainly as you get over, you know, 20, 25 million of ARR stuff is pretty locked down.”
Kyle says that customers hesitate when rolling out AI, and that the bigger enterprises especially are really scared about data security and compliance.
“Especially the bigger enterprises are really scared about the data security and compliance”
said Ghazi Masood ([Un]Churned), John Kaplan (Revenue Builders)
He says LLMs will get commoditised and hopes prices come down. He says that although copycats have appeared, which makes it harder to pick winners, vertical AI companies offer an immediate cost benefit for older industries.
“these vertical AI companies, you know, that are like domain experts in, you know, a specific vertical. I think those are the companies that will survive is my opinion.”
Kaplan says AI companies approached Force Management wanting to be partners because they were experts in technology but lacked sales domain expertise. He argues that customers have judged sellers in the same ways for thousands of years, so the human factors of selling must be built into the tools.
“They were the experts in the technology and what they didn't have was domain expertise or what I call human factor knowledge.”
said Grant Clarke ([Un]Churned), Ian Tickle (The Revenue Leadership Podcast)
Grant Clarke says renewal reps normally start each day behind and triaging inboxes and CRM tasks. In Atlas, the agent researches overnight or in the background, checking each account's last touch and any responses to outreach. It then decides whether it can act autonomously under its guardrails, rules and deterministic settings, or must escalate or hand off to the human. Escalations go to the top of the list, such as a customer saying they have three technical issues and won't renew, and the rep picks them up with context from the agent.
“And what the agent is trying to decide is, can I as the is the autonomous agent handle this based on my guardrails, my rules, the deterministic settings that I have.”
Ian has a Claude Cowork setup that tells him each day whether pipeline is within the tolerance he needs for the business to close. If it is on track, he doesn't need to look or hold a review meeting just to confirm things are fine. He thinks roles will increasingly centre on being alerted when something is off. He also expects frequent small updates from tools to inform decisions that feel instinctive but rest on a week of context.
“I have a Claude Cowork that every day tells me, you know state of my pipeline. And it tells me basically is it within the tolerance to arm happy with for the amount of business that we need to close?”
said Ross Simmonds (The Dave Gerhardt Show), Aviv Canaani (The Revenue Leadership Podcast)
Ross's example is podcast repurposing. Finding key moments used to take hours of listening before the clips were sent to an editor, and AI can now do much of that work. He says AI has created challenges, but he sees the opportunity as massive.
“We now have the ability to distribute great content, valuable stories across multiple platforms for a fraction of the cost that it used to take.”
Aviv says a new video used to need a script, a production company, and around $30,000 to $40,000, and now the team produces videos in days with tools such as Sora. He says the marketing team produced more than 100 campaign pieces in a month, which he could not have done without AI. He gave a Bob Ross Sora video on TikTok that got more than a million impressions as an example.
“it's going to cost you 30, $40,000”
said Simon Farthing ([Un]Churned), Alex Bilmes (Revenue Builders)
He said the initial build took as little as three or four hours, but that there were hundreds of bugs and many plugins, add-ons and connectors added over time, so the total was perhaps 30 or 40 hours. He built it on nights and weekends, starting after a session with AI-focused leaders where Claude ran in the background.
“Maybe 30 or 40 hours kind of working on the mission control”
Alex Bilmes said many AI tools can handle around 80% of a task, and prototyping is easy. He said turning that into something production-worthy for a specific organization is hard, and that this is why leaders are increasingly asking domain-specific vendors for forward-deployed help.
“It's really hard to turn it into something as production worthy that works for your organization.”
said Kyle Norton (The Revenue Leadership Podcast), Manish Chawla ([Un]Churned)
Kyle Norton says he has become more optimistic about Salesforce after the Einstein years, citing its acquisitions and partner and integration ecosystem. He says companies with $100 million or more in ARR are unlikely to rip out Salesforce for an AI CRM, though a 10-person marketing shop might. Aviv agrees, expecting more consolidation into one system and saying he would not drop Salesforce to vibe-code a CRM.
“Nobody's ripping out Salesforce for an AI CRM when they're like 100 million plus ARR.”
Manish said larger customers ask about cross-enterprise AI solutions from providers such as Google. He said PowerSchool's counter is that all the data sits with it, so it can be more effective with AI.
“Our claim to fame is we are the system of record, right. So all the data sits sits with us. Therefore we can be more effective.”
Asked by Mark whether AI use cases that look like features incumbents could capture limit the opportunity for founders, Kyle says there are already too many point solutions layered onto the sales tech stack. He says startups will create much of the innovation, but because existing platforms have the reach, they will probably acquire some of this technology, and something like a virtual AI SDR is probably better integrated into an existing CRM than bought as a point solution.
“So I think we're going to see a lot of acquisitions in the space where maybe the innovation's coming from startups, but given that existing platforms have the reach, they're probably going to be acquiring some of this technology.”
said Jordan Crawford (Topline), John Kaplan (Revenue Builders)
The host says you should not buy technology just to buy it or to have relationships. He says each use case should be matched to a high-value sales workflow that makes sellers more productive, starting with areas such as qualification, and that tools should reach curious sellers seamlessly and with little pain.
“You're not buying technology just to buy it or to have relationships.”
Jordan says that when he works with customers he asks them to define the output first and show the exact thing they want, then identify where the inputs are. Problems are ill-defined for AI when the output is something like a closed-won deal. He suggests asking what you would have engineers build if you were a product manager.
“So when I work with my customers, I say, define the output first. Show me the exact thing.”
Jordan says telling RevOps to make things more AI is too vague to act on. He suggests naming the team and the task, such as an SDR team's work, and removing that task to save them time. He adds that RevOps overlaps closely with what AI can do but needs a strategy and a CRO who knows the tools.
“you need to say, well, okay, our SDR team is doing this, we need to like pull that out of their work and save them time.”
Kaplan says he applied AI to the challenges where the promise was productivity, capacity and speed, looking at what caused him problems in those areas as a seller. He says he did not really learn the technology until he applied it to those problems.
“At least in my case, I really did not learn until I applied it”
said Glenn Fogel (Grit), Sahir Azam (Revenue Builders)
Fogel describes a three-way arrangement among the restaurant, OpenTable and a voice AI company. When his wife calls a restaurant, an AI takes the reservation. He says restaurants don't want to pay staff to answer phones, and callers who prefer a human can still speak to one.
“So it's a three -way type thing with the restaurateur us OpenTable”
Azam said large amounts are spent on call centers, often outsourced overseas, just to collect basic information, and predicted more of these interactions will be driven by voice agents. He gave a bank auto-dialing a customer to check for a fraudulent transaction as an example that can already be automated. He said voice has to feel natural, because a caller will hang up if they clearly hear a machine.
“More and more of those interactions are going to be driven by voice agents across a variety of different areas.”
said Alex Mashrabov (The Twenty Minute VC), Marc Ferrentino (Topline)
Mashrabov cites Higgsfield's open-sourced AI-generated movie project. In it, average prompt length was over 3,000 words and each scene used at least 10 image references to define characters, backgrounds and positioning. He compares video models to a modern rendering engine like Unreal or Unity, which he says cannot be directed through text alone.
“for 90 minutes of, let's say, TV quality content, it was over 100 hours of AI-generated content. So creative decisioning, picking the right piece, is still very important.”
Marc says that despite tools that make video creation look easy on social media, a video he made took him about 27 straight hours to look the way it did. He says it is not a one-shot process and still requires taste and energy. He expects video to keep thriving and wants easier ways for people without artistic ability to produce something that does not look like garbage.
“it still took me like 27 straight hours to get it to look the way it did.”
said Alex Mashrabov (The Twenty Minute VC), AJ Bruno (Topline), Asad Zaman (Topline)
Everyone, including the creative team, moved to Claude from March to June, which is when creatives started vibe coding. Coders moved to Codex as of mid-June, and top creatives followed over time. Mashrabov calls these shifts 'so cyclical'.
“But then we started to see that all the coders quickly moved from Claude to Codex as of mid June.”
He said the switch to Anthropic was quite easy and compared it to moving cloud providers, though easier. He said he is pretty bearish on central AI platforms being sticky right now. He said the switching costs at the model layer are low.
“the move to anthropic was quite easy. It's a bunch of markdown files.”
Asad ran an experiment in which he reduced his OpenAI use and used Google and Anthropic more. He said the product feels sticky while in use but he does not miss it once he moves to another product. He argued OpenAI must solve stickiness to become a great future company, and that it has not solved it yet despite its consumer lead.
“And so there's not actual built-in stickiness there.”
said Rebecca Nerad ([Un]Churned), Mark Vovsi ([Un]Churned), Adnan Rahman ([Un]Churned), Rob Edmondson ([Un]Churned)
Nerad said the better AI captures risk signals, the sooner and more proactively the team can reach out before a customer churns. She described this as something she is still piecing together, and said she can see it happening but has no measurement for it.
“To me, it's the idea of time to response. The fact that we can act on potential churn risk sooner”
He said CS has promised to catch churn early for 10 to 15 years, but he thinks it is now becoming possible because the technology is here. He said teams can pick up signals from telemetry and do analysis in minutes instead of hours. This was his stated expectation for the next 12 months rather than a result he had already measured.
“It's really becoming possible right like you can pick up on signals using telemetry Not three months not nine months a year in advance”
Mark said the first version was a couple of weeks from release and that Proofpoint is partnering with Quadsai on predictive metrics. The system is intended to show which accounts need attention and what to do for each one, and to push results into the CSP and the control tower eventually. He said it currently runs on a separate site and that the goal is to have it live in the CSP.
“we're at structured data, unstructured data, human reported, calculated, machine learning, we're partnering with somebody going to work with predictive metrics.”
Adnan says the model uses a variety of flags and filters to indicate churn risk, and it provides insight into why a customer might churn. He says the next stage is a propensity-to-buy signal, flagging customers who have a need or use case for a product they do not own.
“we just released a churn prediction model tool, right? So it actually tells us, hey, a customer might have a high propensity to churn based off a variety of different flags and filters.”
Rob said Ironclad has a robust view of customer data covering both product usage and interactions across the business. The team worked with a vendor to build the model, which he described as very data intensive. He framed it as a project to predict six months out whether a customer will return.
“we did a project with the vendor to say, let's if we get what can we do to build a model to predict six months out”
said Mark Roberge (Grit), Joubin Mirzadegan (Grit)
Having looked inside some of them, he says many distribute through a PLG model (which he says was invented at Dropbox in 2010) adopted by SMBs and frontline ICs who switch tools freely on experimentation budgets rather than production budgets, so their retention indicators are poor. On go-to-market he sees no cap on burn ratios and unrealistic unit economics in many cases, plus circular revenue issues, and compares some of them to Groupon and WeWork. He grants there is some 'gold in there.'
“taking advantage of an experimentation budget as opposed to a production budget.”
He compared this to COVID-era talk of product-led, bottoms-up growth, where a product sells itself only while demand is insatiable and companies later have to distribute the product themselves. He said he worries less about the large AI players with good customer usage and more about startups at the fringes of cybersecurity trying to automate through AI.
“I do think that at least for these AI native applications and infrastructure companies, the thing that where I think the music will stop is that there is basically an insatiable demand to just continue to consume this technology.”
said Peter Zaffino (Grit), Bret Taylor (Grit), Arvind Jain (Grit)
He said he 100% buys the AI hype and that it is for real. He said that even if time froze and other companies caught up, the technology would still transform businesses across insurance, banking and manufacturing value chains. He said agent orchestration could bring more accuracy, more insight and a speed of execution about 10 times faster.
“even if you froze time now and didn't make the progress, let everybody else catch up, it will still transform businesses.”
Bret Taylor offered this as a thought exercise, saying that many tasks already need less intelligence than current models provide. He said this is his guess and that the implication is interesting in itself.
“if we paused innovation and just absorbed the intelligence of all the existing models, my guess is there's still trillions of dollars of economic value we haven't realized yet”
He says people sometimes get carried away asking whether core model improvements will slow down. He says that question is not relevant in many ways to most business problems today. His point is that far more can be done with current model capabilities.
“Like it's not relevant in many ways to most business problems today.”
He says he would put current usage of model capabilities at not even 1%. He expects that even if models stopped improving, AI-powered products will grow massively across industry verticals over the next five years, with usage moving from 1% to 10% to 20%. He says most business problems need significant work built on top of the models to generate value.
“And I would say that we've not even used 1% of current capabilities of these models, not even 1%.”
said Kyle Norton (The Revenue Leadership Podcast), Jordan Crawford (The Revenue Leadership Podcast)
Moss routes everything through one Telegram channel with a smart routing agent. He said he probably should have used multiple threads, because deep work across many agents in one thread gets confusing, and Grokbot's separate channels were the first thing that drew him to it.
“I separated podcast production from podcast outreach because having them having it go back and forth in the same channel, I found personally confusing”
Kyle Norton says he spent 10 to 15 hours building an XDR hiring analysis, and when he got stuck and kept spinning he moved to a new terminal window and started again, which made progress. He explains that every token above in the conversation is pulled into the context window to answer each request.
“Okay, I just I need to start in a fresh context window.”
Jordan's campaigns each have a folder holding only the transcripts for that campaign, so historical transcripts that do not apply are excluded. He said this makes the context get better over time. Claude Code's ability to work in one folder is what makes this possible.
“that context gets better and better and better and better.”
said Kyle Norton (Topline), AJ (Topline)
Kyle Norton says Steve has an OpenClaw instance with context files, skills and APIs connected to everything. He expects tools like Claude's co-work and dispatch to bring this level of AI use to many more people. He predicts this will show up in GDP numbers over the next 12 to 24 months. Another participant adds that it will expose people who don't adopt it.
“I think the average employee in the next six to 12 months will be much more like a Steve.”
He defined productivity as the number of tickets cleared through QA after a story is complete, and said Claude Code could take productivity per engineer up five times next year. He called it a bold prediction and said engineers remain important.
“you will not have an engineering team that is not three to five X more productive today at the end of 26”
said Josh Schachter ([Un]Churned), Abbas Haider Ali ([Un]Churned)
Schachter said he expects many specialized applications, such as onboarding software, but thinks the CSM role will widen end to end, with the depth of each CSM coming from how they use the full set of specialized AI tools. Richmond responded that one person will be able to do a lot more with copilot-style AI.
“I actually think that the CSM is going to become a little bit more generalized.”
He said functional capabilities are limited by the knowledge any one person can retain. Large language models are good at knowledge retrieval and combining material across domains. He applies the same idea to his personal knowledge, saying he can only actively retain so much even about topics he has written about.
“a lot of our functional capabilities are limited by the knowledge any one person can retain”
Abbas described CSM skills as radar charts covering commercial, relationship, tactical depth, domain knowledge and portfolio management, with the specialized generalist being specialized in using AI tools. Josh summarized this as a T-chart whose vertical is AI tooling, and said AI will give more breadth across functions but people need to become experts in how to use it.
“It's AI is going to give us more breadth, but we will need to really become experts in how to use it for all those different pieces of our work.”
He said the shift to 7% is driven by a generalist CSM who can handle support, services and technical questions with AI tools, so one person covers more customer needs. Specialists are then freed from repeated problems to move up the value chain. He framed this as a prediction for how the role will evolve.
“ultimately, it's this sort of thing that gets you from that 10% to 7% because you can cover more ground.”
said Jonathan Moss (The Revenue Leadership Podcast), Kyle Norton (The Revenue Leadership Podcast)
He says you may outgrow the platform, get in too deep, or later have the expertise to build it yourself, so ease of moving off it matters most. Norton said platforms are acceptable as long as you can see the intelligence inside them and take it out, because the brain is a compounding asset.
“the portability is the most important factor if you're going to do it”
Kyle said he gets nervous when a platform tells reps what to do next without showing why. He prefers tools where he can inspect markdown files and context, and said he left Gong over its walled garden for Momentum, which pushes context into Salesforce and his own prompts. Adrian agreed that locking intelligence in is not a moat worth having, and said he works hard to keep his Gong calls out of Gong. He expects buyers and agents to pick tools that give agents better context.
“I don't really want to use tools that trap my intelligence.”
said Simon Farthing ([Un]Churned), Adnan Rahman ([Un]Churned)
Simon described agents that would sit on a customer's shoulder, flag when they have just done something that may not be what they wanted, check they have not broken anything, and point them to unused parts of the platform. He said he expects these agents to have a disproportionately strong impact with smaller customers, though this is still experimental.
“which will basically sit on your shoulder as a customer and say, hey, you just did something.”
Adnan says an agent could reach out to customers when a need appears to gauge interest and then create a lead. He also points to renewals, including month-to-month customers affected by price increases, where he says outbound at Paycor's scale is hard. He frames these as his magic-wand priorities rather than deployed work.
“So obviously, the renewals and expansions, I think are two big ones.”
Actions written 10 Oct 2026 from the most useful of 636 recent insights and checked against them.
In a mock investment committee on ElevenLabs at $22B, Jason Lemkin argued for putting 10% of the fund in. He said ElevenLabs' lead keeps widening, its margins are surprisingly strong and it could approach $1B in revenue. He had earlier predicted cost-driven substitution in voice, but now said a phone agent (e.g. for a flower shop) must answer in seconds and get the order right, which limits price pressure. He was shocked that relatively modest companies happily pay hundreds of thousands of dollars a year, though he said 12 months out is hard to predict.
“I'm shocked at the number of folks that are relatively modest companies paying hundreds and hundreds of thousands of dollars a year happily.”
Rory O'Driscoll said he would not guess 50% for US open-weight share, but that even 10–20% of tokens would be extremely significant. He reasoned that this traction would come from the more enterprise-centric, safety-conscious customers. He noted it would be risky for inference providers.
“the foundation models on OpenRouter I think get a 20 30 percent of tokens and get 90 percent of the dollars.”
Dev Ittycheria said enterprises will care a lot about whether they can trust these systems. Following a recent Washington conference, he expects a self-regulating approach to emerge. He said the hard part is defining safety, and that a benchmark-like mechanism is needed even though benchmarks can be gamed.
“The more safe these models are with whatever mechanism that plays to prevent people from doing bad things, the more widely usable they'll become.”
Asked what share of open-source tokens will run through US versus Chinese models in 12 months, Jason Lemkin said he could see it being half, given developer willingness to switch and the importance of sovereignty. He explicitly said he was not predicting it would happen, because he is skeptical whether parity is real or 'pretend'.
“But I could see it being half in 12 months.”
Having recently swapped models himself, Jason Lemkin said you have to requalify prompts and redo workflows, contrary to what 'the internet says'. He still called it much easier than swapping out a database.
“You have to qualify your prompt. You have to redo your workflows.”
Jason Lemkin said any eval can be made to look great. What matters to him is speed, real outputs, and performance on real and mission-critical workflows, such as correctly stating whether an item is in stock. He said evals focus on things like 'chess championships', so he would want to try Beam himself, for example on OpenRouter, before believing the claims.
“I found that every single eval needs like three asterisks and four daggers next to”
Jason Lemkin said everyone he spoke to at Dreamforce was overloaded with internal demand for tokens. He said this may lead to suboptimal model use and waste, but everyone has to manage it. Combined with reluctance to use Chinese models, he sees it as a big opening for US open-weight providers.
“the demand is probably an order of magnitude more than they've figured out how to surface”
Dev Ittycheria said there is a big gap between AI capabilities and enterprise skill sets. As costs drop and staff become more conversant, companies will deploy AI for more and more use cases. He said he sees this at MongoDB, and that the market is massive enough for both frontier labs and open-source providers to succeed.
“I think we're going to see Jevons Paradox kind of come into play”
Rory O'Driscoll said enterprises have so far defaulted to frontier models because people are lazy and bills were not astronomical. If OpenAI and Anthropic each reach about $70B in ARR, that $140B becomes a meaningful chunk of US corporate profits. He expects CFOs to push for cheaper models on non-frontier tasks, keeping Anthropic for the hardest problems, and to slowly take 'slugs of revenue' from the frontier labs.
“We need to shave three million off this token bill.”
Dev Ittycheria said a model within about six months of the frontier that covers 90% of use cases would be attractive, especially for reasoning workloads like coding and agents. He said he didn't see why enterprises would not want to talk to such vendors immediately.
“this also reinforces the point that you don't need the frontier level intelligence for every workload.”
Jason Lemkin said he had a surprisingly large number of executive conversations at Dreamforce and nobody really wanted a Chinese-sourced model, 'right or wrong, fair or not'. Those using them felt they were doing so under duress for cost. He said a US model that 'nails' it could see 'a torrent of demand'.
“nobody I talked to really wanted to use a Chinese source model, right or wrong, fair or not.”
Dev Ittycheria (disclosing that Sequoia is an investor in Reflection) said Reflection's Beam gives enterprises near-frontier intelligence at three to four times lower cost. He said enterprises always hesitate over Chinese open-source models, especially in regulated industries or ones handling sensitive data. He called the launch very noteworthy for the enterprise.
“One, you get a US open source model that's near frontier intelligence. That's three to four times more cost effective, i .e. cheaper.”
Dev Ittycheria said he had just been discussing token costs and token budgets with MongoDB's CIO, noting no enterprise has unlimited budgets. The second factor he named was IP rights: how models are trained, who gets the data and how proprietary it is. He said these two are pretty important points when it comes to which models you choose and who you work with.
“I was just talking to our CIO and we're talking about token costs and token budgets, right?”
He says incoming client data is always an issue and a common slow point. Agentic tools sit on top of the implementation process to streamline it and find data errors. He gives this as one reason implementations are faster and less expensive than in the past.
“one of the reasons our implementations are faster and not as expensive as in the past. We have a lot of agentic tools sitting on top of that process to streamline it, to find errors and data is always an issue”
The first leg is configurable models (seasonality, elasticity, and any dynamic attribute such as inventory) that set prices, plus AI for matching items in scraped competitor data and flagging which items to promote. The second is internal agentic tools used for the product roadmap, QA and bug fixes, and implementation. The third is natural-language agentic capabilities in the product UI, which he says they are enhancing continually.
“So that's kind of the core of almost like a three-legged stool is the way I explain it to our clients”
Ross calls Reddit 'ridiculously valuable' still. He says people don't realize how much B2B dealmaking it shapes, because it is such a powerful source of training data for AI tools like Gemini. Manipulating and spamming Reddit is the 2026 marketing trend he hopes ends.
“There's billions of dollars worth of deals in B2B that are being influenced by Reddit that people don't even realize because Reddit is such a powerful training data, platform.”
Ross says Foundation created the piece for a client about three years ago. Instead of publishing it once, they kept updating, distributing and promoting it. It now appears in LLM answers to some of the highest-priority prompts in the client's category, and the client still generates leads from it.
“We continue to update it, distribute it and promote it is now cited in the LLM consistently across some of the most priority prompts in their category, and they're generating leads on the back of this asset from, like, three years ago.”
Ross says LLMs pull from Reddit, SEO and the other channels that organic distribution targets when they answer questions. He says demand for Foundation Marketing's services has increased as a result. He also launched Distribution.ai as software for marketers and sales executives to distribute their stories at scale.
“Ai is now using all of those different distribution surfaces to translate back into the LLM answers to questions. So our services have increased in demand because of that.”
Ross's example is podcast repurposing. Finding key moments used to take hours of listening before the clips were sent to an editor, and AI can now do much of that work. He says AI has created challenges, but he sees the opportunity as massive.
“We now have the ability to distribute great content, valuable stories across multiple platforms for a fraction of the cost that it used to take.”
Ross says low-quality AI content is not limited to Reddit but fills X, Threads, LinkedIn and Quora. He argues that brands that put authentic, valuable human content on these crowded platforms can win.
“That creates an opportunity If you can create authentic, valuable human content and then put it on these platforms, you can win.”
Vercel grew from about 600 to just over 800 people in her roughly year and a half there, while revenue grew 'well north of triple digits'. She said she picks companies where revenue grows roughly exponentially while headcount grows roughly linearly, as at Google and Stripe. She attributes part of the difference to AI giving many roles multiple-times productivity.
“there are a bunch of things where like, you know, the equivalent human is getting NX productivity.”
He described reps running meeting notes from Granola through Claude and sending the output to the team. He also said every AI-made deck now looks the same. He told his team that if he doesn't want to read these, customers definitely don't. Grosser added her own complaint: AI-written call prep docs instruct her in an unnatural tone, such as telling her not to sell on an exec-to-exec call.
“your job is to synthesize what actually matters.”
Grosser described analysing call transcripts for patterns as a way to get product and go-to-market to agree on what is and isn't working. Objections that can't be handled because the product lacks something are automatically submitted as product asks in Vercel's internal GTM feedback tool.
“we proactively extract objections from calls and score the degree to which we've handled them.”
Grosser said Vercel has a data science agent you can ask anything you would have sent to RevOps. As a result, she described the RevOps team as leaner and more efficient than teams she had in prior roles.
“We have a great data science agent that, like, I mean, you can ask it anything you would have sent to RevOps.”
Grosser said Vercel gets a lot of support volume and 91% of it is resolved by an agent. The remaining support team, which is technical, now focuses on legitimate product problems, works closely with engineering, and in some cases writes pull requests to fix issues itself.
“Vercel gets a lot of support volume, as you'd expect, and 91% of it, we resolve with an agent.”
The investor wasn't using specialized tools, just basic workflows while switching between Claude and ChatGPT. He now spends 25% of his former time on the same work, freeing 75% for other things. The speaker presented it as an example of how well AI works where it fits.
“He's like, my job is now 25% of what it used to be.”
The suggestion is that organizations need a controlled sandbox to design, test and stress an end-to-end agentic version of an operation before deploying it as the main process. Without that, the speaker expects companies to look only slightly different seven or eight years later. A host said this pacing should be reassuring compared with predictions that everything changes within a year.
“Till that happens, the pace at which enterprise will move is the same pace we've always seen them to move, which is like seven, eight years later”
A host asked whether the clearest form of reimagining is to sell the outcome rather than the form-filling work, citing how factories were redesigned around electricity. Manny agreed. Another host added that chat interfaces should give way to software that already knows what to do.
“At the end of the day, the process exists for a reason. And that reason is the outcome. And all the middle steps are all negotiable.”
A host argued that even frontier models fail on 3% to 5% of multi-step tasks, which limits what enterprises can trust AI to do, even with forward-deployed engineers on top. Manny replied that teams had the same problem before and that AI isn't a magic bullet. Frequent misfires usually point to a harness, memory or context problem that you debug until it's fixed. He says the goal is an organization that needs fewer people, has more throughput or delivers higher quality.
“when you're seeing a lot of errors or a lot of misfires, then you may have a harness problem, maybe you have a memory problem, maybe you have a context problem”
Manny says software vendors used to treat professional services as taboo. He is excited by a new breed of companies that sell software together with transformation of the customer's operations, rather than layering AI onto existing workflows.
“there's this whole new breed of companies who are like, who are selling software with transformation.”