“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.”
AI
Where they agree
-
Reliable AI workflows are mostly deterministic code and rules, with the LLM used only at narrow interpretive steps and never for math.
6 independent voices · 3 shows3 new this month
On Topline, Tim Rutten said real bank AI deployments are almost 99% deterministic, with the language model reasoning at only one heavily guardrailed moment.
8 sources
AI pricing becomes trustworthy when it uses deterministic inputs, formulas and math rather than guessing or searching the web. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Keith Peiris sees the LLM's role in forecasting as surfacing insight, such as which deals were on the cusp because of missing features or competition, not doing the arithmetic. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
LLMs shouldn't do math, and converting forecast formulas from prompts to real code got Lightfield's dashboards about 95% of the way there. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Real bank AI deployments are almost 99% deterministic, with the language model reasoning at only one small moment. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Rutten estimates that roughly 70 to 80% of GTMOS is deterministic, with LLM calls used only for specific activities. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Teams too often start with AI and look for a problem, when they should start from the business challenge and use deterministic flows where they can. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Judgment calls such as tone analysis are not something Jordan gives to AI, while deterministic steps can be broken out for it. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
The agent workflows that work today combine deterministic nodes with an LLM or agentic step in the middle. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
-
Centrally built, top-down AI systems deliver business results, while giving everyone tools to build their own does not.
6 independent voices · 3 shows
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.
9 sources
Justin favours a centralized AI model with role-based playgrounds for each function, so revenue answers stay consistent and token use stays efficient. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Decentralized AI adoption can lead to conflicting data and derail leadership meetings. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
The key to scaling is a shared landing zone with enterprise-level authentication, connectors and guardrails, rather than individual AI instances. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Centralization matters more than individual AI use for keeping messaging consistent across a sales organization. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
A rep pitched new SKUs and a product offering that did not exist, because ChatGPT suggested it. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI sourcing should be centralised in RevOps and delivered to SDRs as prepared data, rather than given to reps as a research tool. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Deploying AI through one central team, built into the tools reps already use, reduces change-management friction. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Personal AI use gives faster emails, notes and decks, but not business use cases that move the needle. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Decentralized AI rollouts, where every employee gets AI accounts and builds their own tools, did not produce business results at Owner.com; centrally built systems did. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
-
AI output must be checked by people with domain expertise, because it is persuasive and certain rather than accurate.
7 independent voices · 3 shows3 new this month
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.
7 sources
Good master data is fundamental to making AI work in pricing, and AI still needs human commercial judgment on top. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Experienced pricers can train AI better because they know what good looks like from trying and failing. Listen
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?”
Listen to the episode Episode AI Link to this Report a problem
People trust AI more on topics they do not know, even though they can see its errors on topics they do know. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI's persuasiveness and certainty are not accuracy, so subject-matter experts must challenge outputs. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
She worries about AI-built plans when teams rely on the AI alone without human review. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI output should never be taken at face value, and that the human in the loop provides curation and context. Listen
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?”
Listen to the episode Episode AI Link to this Report a problem
Generative AI layered on complex data models can cause serious errors when users trust the bot's output without checking it. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
-
Companies should buy AI tools rather than build them, because keeping in-house builds current as models change costs too much.
5 independent voices · 4 shows
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.
5 sources
Building AI tools in-house carries ongoing costs of maintenance, versioning, iteration and after-hours outages that are often left out of the decision. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Building and maintaining an in-house AI workflow would consume far more resources and become outdated in about six months. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Deltek built an AI support tool while evaluating buying in parallel, and chose to buy because buying avoided the upkeep needed as models changed. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Rebuilding CPQ in-house with AI requires believing a long chain of hard things all go right. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Richmond predicts the big AI winners over the next few years will be application companies, not businesses building their own tools. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
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.
Where they split
Centrally built, top-down AI systems deliver business results, while giving everyone tools to build their own does not.
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)
9 sources
Justin favours a centralized AI model with role-based playgrounds for each function, so revenue answers stay consistent and token use stays efficient. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Decentralized AI adoption can lead to conflicting data and derail leadership meetings. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
The key to scaling is a shared landing zone with enterprise-level authentication, connectors and guardrails, rather than individual AI instances. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Centralization matters more than individual AI use for keeping messaging consistent across a sales organization. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
A rep pitched new SKUs and a product offering that did not exist, because ChatGPT suggested it. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI sourcing should be centralised in RevOps and delivered to SDRs as prepared data, rather than given to reps as a research tool. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Deploying AI through one central team, built into the tools reps already use, reduces change-management friction. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Personal AI use gives faster emails, notes and decks, but not business use cases that move the needle. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Decentralized AI rollouts, where every employee gets AI accounts and builds their own tools, did not produce business results at Owner.com; centrally built systems did. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI adoption should start bottom-up with individuals and organic champions, because restricting or forcing it backfires.
said Jonathan Moss (The Revenue Leadership Podcast), Jared Collins ([Un]Churned), Guy Galon ([Un]Churned), Josh Schachter ([Un]Churned)
4 sources
Experity first restricted AI tools to a select group, which angered people and pushed them to outside tools; it then gave everyone access to connected tools to curb sprawl. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
AI adoption should start with individuals automating day-to-day repetitive tasks rather than enterprise-wide initiatives. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Guy tells his team to set small AI use cases rather than big targets, so that the next steps come from the team. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Josh advises finding the person ahead on AI in the organisation and championing them, because he says adoption cannot always be forced. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Centralists focus on production systems that touch shared revenue data and security, while bottom-up advocates focus on building individual habits and buy-in.
Companies should buy AI tools rather than build them, because keeping in-house builds current as models change costs too much.
said Rob Potter (Revenue Builders), Snehal Nimje (Topline), Margo Martin ([Un]Churned), Joubin Mirzadegan (Grit), Jim Richmond ([Un]Churned)
5 sources
Building AI tools in-house carries ongoing costs of maintenance, versioning, iteration and after-hours outages that are often left out of the decision. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Building and maintaining an in-house AI workflow would consume far more resources and become outdated in about six months. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Deltek built an AI support tool while evaluating buying in parallel, and chose to buy because buying avoided the upkeep needed as models changed. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Rebuilding CPQ in-house with AI requires believing a long chain of hard things all go right. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Richmond predicts the big AI winners over the next few years will be application companies, not businesses building their own tools. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
A company's core AI intelligence and context should be built and owned in-house rather than bought.
said Jonathan Moss (The Revenue Leadership Podcast), Tim Rutten (The Revenue Leadership Podcast), Kyle Norton (Topline)
3 sources
Moss stopped Experity from putting its business logic and company context into a third-party product, arguing the brain is proprietary and must be built in-house. Listen
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?”
Listen to the episode Episode AI Link to this Report a problem
A bespoke signal engine with a rubric the team tunes themselves can replace a commercial tool like 6sense. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Own the core intelligence internally and buy workflow and user-experience tools, according to Kyle Norton. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
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.
AI's value comes from redeploying freed time into higher-value or new work, not from cutting headcount.
said Ian Tickle (The Revenue Leadership Podcast), Parag Agarwal (Grit), Chael Banks ([Un]Churned), Brad Casemore ([Un]Churned), Jared Collins ([Un]Churned)
6 sources
Efficiency from AI and tools is wasted unless the freed time is redeployed into more valuable work. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Agarwal predicts the next wave of AI value comes from entirely new work rather than doing existing work faster or cheaper. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
The real business value of AI agents is people expanding their span of ownership, not doing the same work faster. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
The goal of AI at Okta is to make people more effective, not just more efficient. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Brad Casemore describes a renewal agent at PartsSource that prepares renewals and reads customer sentiment, freeing the renewal team for proactive outreach. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Collins expects AI to surface patterns and problems that teams did not know existed, beyond making known tasks faster. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI lets teams deliver the same work with substantially fewer people.
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)
5 sources
AI has made Vercel's hypergrowth feel digestible, unlike when Stripe doubled headcount in a year during COVID. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Justworks never filled a planned onboarding coordinator role because AI now handles new-hire scheduling, knowledge checks and manager alerts. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Slido reduced its customer voice program from four full-time employees to about 1.5 using AI. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Richmond wants AI-enabled CSMs to cover more accounts efficiently rather than broaden their scope. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
AI SDR tools mean Box doesn't need the same SDR capacity, which raises the question of whether to bank the savings or fund other initiatives. Listen
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?”
Listen to the episode Episode AI Link to this Report a problem
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.
AI has collapsed the cost of producing and distributing marketing content.
said Ross Simmonds (The Dave Gerhardt Show), Aviv Canaani (The Revenue Leadership Podcast)
2 sources
AI has cut the cost of distributing content across multiple platforms to a fraction of what it was. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI tools now let DataRails make far more campaign content, with production that once cost $30,000 to $40,000 per video Listen
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”
Listen to the episode Episode AI Link to this Report a problem
High-quality AI-made content still demands substantial human taste, selection and effort.
said Alex Mashrabov (The Twenty Minute VC), Marc Ferrentino (Topline)
2 sources
Producing 90 minutes of TV-quality AI video took over 100 hours of generated footage, so human creative selection still matters. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Making a polished video still takes taste and substantial effort, even with new tools. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
They differ on quality bar: volume repurposing and campaign content are cheap now, while polished creative output still needs heavy human effort.
From one operator's experience
What one named guest described doing or seeing. Each is a single account, not a point several operators agree on.
“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?”
What to do
- Design agent workflows as fixed steps with one guarded LLM decision in the middle, and move any calculation into code, as Wade Foster, Tim Rutten and Keith Peiris (Topline) describe.
3 sources
The agent workflows that work today combine deterministic nodes with an LLM or agentic step in the middle. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Real bank AI deployments are almost 99% deterministic, with the language model reasoning at only one small moment. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
LLMs shouldn't do math, and converting forecast formulas from prompts to real code got Lightfield's dashboards about 95% of the way there. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Before deploying multiple agents, build a company dictionary for terms like champion and ARR and a single approved source of truth, following Alex Bilmes (Revenue Builders).
3 sources
Defining each company's own terms, such as champion and ARR, in a semantic dictionary keeps AI answers consistent. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Multiple AI agents pulling from different sources give inconsistent answers to the same question. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
A rep pitched new SKUs and a product offering that did not exist, because ChatGPT suggested it. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Measure the share of each rep's week spent with customers before and after AI rollout, as Mark Roberge Roberge ([Un]Churned) urges boards to demand.
2 sources
Selling time is the revenue velocity variable AI can most realistically double, since best-in-class teams have historically spent about 30% of their time selling. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Selling time, the share of a week a rep spends with customers or prospects, has historically been 25 to 30 percent, and could reach 75 percent for the best AI-enabled teams. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Require users to correct bad AI outputs inside the system instead of reverting to old methods, as Jonathan Moss did at Experity (Revenue Leadership Podcast).
2 sources
Moss required users to correct bad AI outputs inside the system rather than work around them, so each correction feeds the brain and the mistake doesn't recur. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Webflow's email agent drafts customer notes from Slack requests, call history and calendar in Adrian's voice, and close to 90% go out unedited. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Compare AI-captured CRM fields against call transcripts before letting AI write them automatically, then back-test prompts on past calls to judge changes such as new pricing, as Kyle Norton (Revenue Leadership Podcast; Topline) did.
2 sources
Kyle's team uses Momentum to write CRM fields to Salesforce directly, after finding its accuracy better than rep-entered data Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Use AI to scan every sales call and fill in CRM fields, then back-test prompts on past calls to check new answers and pricing. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
6 more
- Run a central AI team that ships into tools reps already use, while giving everyone governed access through one enterprise-authenticated landing zone, combining Kyle Norton, Jonathan Moss and Tim Rutten (Revenue Leadership Podcast).
4 sources
Deploying AI through one central team, built into the tools reps already use, reduces change-management friction. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Decentralized AI rollouts, where every employee gets AI accounts and builds their own tools, did not produce business results at Owner.com; centrally built systems did. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Experity first restricted AI tools to a select group, which angered people and pushed them to outside tools; it then gave everyone access to connected tools to curb sprawl. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
The key to scaling is a shared landing zone with enterprise-level authentication, connectors and guardrails, rather than individual AI instances. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
- Before switching models to save money, test cost per task rather than per-token price: Jason Lemkin (20VC) found a cheaper model used about 43% more tokens, and swapping meant requalifying prompts.
3 sources
Jason Lemkin's own test of a new model version found it used far more tokens, so total cost fell only about 10% despite cheaper per-token pricing. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Jason Lemkin found that switching LLM providers is real work, even if easier than swapping a database. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Rory O'Driscoll predicts CFOs will force CIOs to cut token bills by routing non-frontier work to cheaper models. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Price in the maintenance of any homegrown AI tool, including versioning and 3 a.m. outages, and default to buying once more than a handful of people use it, per Rob Potter (Revenue Builders) and Simon Farthing ([Un]Churned).
3 sources
Building AI tools in-house carries ongoing costs of maintenance, versioning, iteration and after-hours outages that are often left out of the decision. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Simon's rule of thumb is to build an AI tool yourself when it is for individual use, and to go with a vendor once more than a handful of people use it. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Deltek built an AI support tool while evaluating buying in parallel, and chose to buy because buying avoided the upkeep needed as models changed. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Have each GTM leader bring their one or two most painful processes on video to a QBR and prioritize which to automate, as Monika Saha ([Un]Churned) does.
2 sources
A QBR exercise that finds AI-ready bottlenecks: each leader brings their most painful processes and the group reviews them on video. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Sam Jacobs suggests a hackathon framework: list your job's activities, map them to tasks an AI agent could handle, then see how far you get building it. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
- Tie every AI budget to one specific business purpose, as Peter Zaffino (Grit) did with AIG's underwriting workflow.
2 sources
AI spending should be tied to a very specific business or functional purpose. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Asad Zaman proposes judging AI investments by asking what you would spend $1M on in the next 12 months, such as an AI that runs complete sales cycles for smaller deals. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
- Check early whether the buyer has an AI committee or needs an AI addendum, because Sam Costello (Revenue Builders) warns these can slip committed deals at quarter end.
1 source
AI committees and AI addenda in buying processes can slip deals that looked committed, so reps need to track how each buyer's process has changed Listen
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”
Listen to the episode Episode AI Link to this Report a problem
All 64 positions best supported first
- Reliable AI workflows are mostly deterministic code and rules, with the LLM used only at narrow interpretive steps and never for math.
6 independent voices · 3 shows3 new this month
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)
8 sources
- Centrally built, top-down AI systems deliver business results, while giving everyone tools to build their own does not.
6 independent voices · 3 shows
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)
9 sources
- AI output must be checked by people with domain expertise, because it is persuasive and certain rather than accurate.
7 independent voices · 3 shows3 new this month
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
7 sources
- Companies should buy AI tools rather than build them, because keeping in-house builds current as models change costs too much.
5 independent voices · 4 shows
said Rob Potter (Revenue Builders), Snehal Nimje (Topline), Margo Martin ([Un]Churned), Joubin Mirzadegan (Grit), Jim Richmond ([Un]Churned)
5 sources
- Large, complex enterprise deals will stay human-led rather than be sold by AI agents.
6 independent voices · 3 shows2 new this month
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)
8 sources
- Unmanaged AI tool sprawl causes overlap and loss of control, so a central team must own governance.
6 independent voices · 3 shows
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)
7 sources
- AI agents need ongoing coaching and corrections fed back into the system, like new employees.
6 independent voices · 4 shows1 new this month
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)
7 sources
- Repeatable work should be codified as reusable skill files that any agent can use.
6 independent voices · 3 shows
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)
7 sources
- Companies must fix their data foundations before AI can deliver results.
8 independent voices · 3 shows
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
9 sources
- AI's biggest near-term sales gain is removing admin work to sharply raise reps' selling time.
5 independent voices · 4 shows
said Mark Thurmond (Revenue Builders), Mark Roberge ([Un]Churned, Topline), Sam Jacobs (Topline), Kellie Snyder ([Un]Churned), Frederic Kerrest (The Science of Scaling)
7 sources
- AI analysis of every call transcript replaces manual call review for coaching and insight.
5 independent voices · 4 shows2 new this month
said Jeanne DeWitt Grosser (Grit), Rebecca Nerad ([Un]Churned), Bob London ([Un]Churned), Jason Forget (Revenue Builders), Kyle Norton (Topline)
5 sources
- Most enterprise AI workloads do not need frontier models and will move to cheaper models, with frontier models kept for the hardest tasks.
5 independent voices · 3 shows4 new this month
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)
7 sources
- AI's value comes from redeploying freed time into higher-value or new work, not from cutting headcount.
5 independent voices · 3 shows
said Ian Tickle (The Revenue Leadership Podcast), Parag Agarwal (Grit), Chael Banks ([Un]Churned), Brad Casemore ([Un]Churned), Jared Collins ([Un]Churned)
6 sources
- Real AI gains come from redesigning workflows around outcomes, not adding AI to existing processes.
5 independent voices · 2 shows4 new this month
said Manny Medina (Topline), Ramin Heydari (Topline), Jonathan Moss (The Revenue Leadership Podcast), Mark Roberge (Topline), Pablo Dominguez (Topline)
8 sources
- Humans should approve AI outputs before they reach customers or production systems.
4 independent voices · 3 shows1 new this month
said Keith Peiris (Topline), Jaleh Rezaei (Topline), Pradeep Raman ([Un]Churned), Jordan Crawford (The Revenue Leadership Podcast)
4 sources
- AI adoption should start bottom-up with individuals and organic champions, because restricting or forcing it backfires.
4 independent voices · 2 shows
said Jonathan Moss (The Revenue Leadership Podcast), Jared Collins ([Un]Churned), Guy Galon ([Un]Churned), Josh Schachter ([Un]Churned)
4 sources
- AI already cuts many research, analysis and writing tasks to a fraction of their former time, often a quarter or less.
4 independent voices · 2 shows2 new this month
said Ann Davis (Revenue Builders), Bryan Murphy (Topline), John Kaplan (Revenue Builders), Asad Zaman (Topline)
5 sources
- AI-generated content is generic slop, so authentic human-written content now stands out.
5 independent voices · 4 shows3 new this month
said Ross Simmonds (The Dave Gerhardt Show), Joubin Mirzadegan (Grit), Dave Gerhardt (The Dave Gerhardt Show), Sam Jacobs (Topline), Darren McKee ([Un]Churned)
5 sources
- AI lets teams deliver the same work with substantially fewer people.
5 independent voices · 4 shows1 new this month
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)
5 sources
- Agents succeed when work is broken into small, atomic tasks handled by narrow, modular agents rather than one agent doing everything.
5 independent voices · 2 shows
said Amanda Kahlow (Topline), Gaurav Agarwal (Topline), Kyle Norton (The Revenue Leadership Podcast), Pablo Dominguez (Topline), Wade Foster (Topline)
7 sources
- CRM data captured by AI from calls is more complete and accurate than data entered by reps.
3 independent voices · 2 shows
said Alex Bilmes (Revenue Builders), Aviv Canaani (The Revenue Leadership Podcast), Kyle Norton (The Revenue Leadership Podcast)
3 sources
- Pointing a general LLM at CRM or business systems produces confident wrong answers, so a purpose-built harness or data layer is required.
4 independent voices · 3 shows
said Kyle Norton (The Revenue Leadership Podcast), Mike Carpenter (Topline), Justin Shriber (Topline), Alex Bilmes (Revenue Builders)
11 sources
- Build AI tools yourself only for personal or throwaway use, and buy vendor systems for anything shared, recurring or needing stability.
4 independent voices · 3 shows
said Simon Farthing ([Un]Churned), Adrian Rosenkranz (The Revenue Leadership Podcast), Kyle Norton (Topline), Jordan Crawford (The Revenue Leadership Podcast)
5 sources
- Outside engineering, and especially in go-to-market, AI has so far delivered little measurable impact on performance.
5 independent voices · 3 shows
said Asad Zaman (Topline), Mark Roberge (Topline), AJ Bruno (Topline), Joubin Mirzadegan (Grit), Steve Cox (Topline), Jeremey Donovan (The Revenue Leadership Podcast)
6 sources
- AI should be used to raise the quality of a person's own writing and thinking, with the author still owning the work.
5 independent voices · 3 shows
said Tomasz Tunguz (Topline), Rick Smolen (Topline), Marcy Stoudt (Revenue Builders), Sam Jacobs (Topline), Arvind Jain (Grit)
5 sources
- Go-to-market teams will follow software engineering in adopting AI practices such as harnesses, shared context and domain-built tools.
3 independent voices · 3 shows
said Adam Liska (Topline), Kyle Norton (The Revenue Leadership Podcast), John McMahon (Revenue Builders)
3 sources
- Non-engineers and small teams can now build useful internal agents and apps quickly, sometimes in a day or a weekend.
3 independent voices · 3 shows
said Ghazi Masood (The Revenue Leadership Podcast), Alex Varel (Revenue Builders), Christina Meng ([Un]Churned)
3 sources
- AI-generated meeting prep cuts manual preparation work sharply and leaves reps better prepared.
5 independent voices · 3 shows
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)
5 sources
- Senior leaders benefit from a personal AI chief-of-staff agent that tracks meetings, priorities and follow-ups.
4 independent voices · 2 shows
said Simon Farthing ([Un]Churned), Parag Agarwal (Grit), Jason Goldsmith ([Un]Churned), Margo Martin ([Un]Churned)
4 sources
- Leaders drive AI adoption best by personally using and building with AI in front of their teams.
3 independent voices · 3 shows
said Jason Goldsmith ([Un]Churned), Kyle Norton (The Revenue Leadership Podcast), Josh Schachter ([Un]Churned)
5 sources
- Leaders should let teams spend heavily on AI tokens rather than cap usage.
3 independent voices · 2 shows1 new this month
said Alex Mashrabov (The Twenty Minute VC), Joubin Mirzadegan (Grit), Parag Agarwal (Grit)
3 sources
- A company's core AI intelligence and context should be built and owned in-house rather than bought.
3 independent voices · 2 shows
said Jonathan Moss (The Revenue Leadership Podcast), Tim Rutten (The Revenue Leadership Podcast), Kyle Norton (Topline)
3 sources
- AI agents can now resolve half or more of customer support volume.
3 independent voices · 2 shows1 new this month
said Jeanne DeWitt Grosser (Grit), Aman Narang (Grit), Diego Ballona ([Un]Churned)
3 sources
- AI earns user trust by showing its reasoning and flagging uncertainty rather than presenting bare answers.
3 independent voices · 2 shows1 new this month
said Mark Stiving (Impact Pricing), Mike Carpenter (Topline), Pablo Dominguez (Topline)
4 sources
- Let employees experiment freely with AI, then harvest the best builds and consolidate them into a standard.
4 independent voices · 2 shows
said Saumyo Mukherjee ([Un]Churned), Kellie Snyder ([Un]Churned), Chael Banks ([Un]Churned), Jeremey Donovan (The Revenue Leadership Podcast)
5 sources
- Natural-language data agents let leaders self-serve analysis instead of asking analysts or RevOps.
4 independent voices · 2 shows2 new this month
said Jeanne DeWitt Grosser (Grit), Christina Parra ([Un]Churned), Mark Vovsi ([Un]Churned), Manish Chawla ([Un]Churned)
4 sources
- B2B brands must work to be cited in LLM answers, because buyers increasingly start their research there.
2 independent voices · 2 shows3 new this month
said Ross Simmonds (The Dave Gerhardt Show), Eric Gilpin ([Un]Churned)
6 sources
- Software should be judged by how well AI agents can use it, and agent preference will decide which vendors win.
2 independent voices · 2 shows1 new this month
said Asad Zaman (Topline), Christopher O'Donnell (The Science of Scaling)
2 sources
- Enterprises will increasingly post-train and run their own open-weight models for cost, privacy and control.
3 independent voices · 3 shows1 new this month
said Jack Altman (The Twenty Minute VC), Jake Saper ([Un]Churned), Tim Rutten (Topline)
3 sources
- Published model benchmarks are gamed and say little; only testing on real workflows can be trusted.
3 independent voices · 2 shows2 new this month
said Jason Lemkin (The Twenty Minute VC), Alex Mashrabov (The Twenty Minute VC), Anastasios Angelopoulos (Grit)
3 sources
- AI amplifies existing skill, widening the gap between strong and weak performers rather than lifting everyone.
3 independent voices · 2 shows1 new this month
said Jen Igartua (Topline), Christopher O'Donnell (The Science of Scaling), Rick Smolen (Topline)
4 sources
- Human roles are shifting from doing the work to managing and directing AI agents.
3 independent voices · 2 shows
said Gaurav Agarwal (Topline), Cassie Vaughn ([Un]Churned), Brett Queener ([Un]Churned)
3 sources
- Individuals should commit fixed, regular time to hands-on AI learning and building.
3 independent voices · 3 shows
said Marcy Stoudt (Revenue Builders), Kristi Faltorusso ([Un]Churned), Mark Roberge (The Science of Scaling)
3 sources
- Agent error rates are manageable and need not limit what agents are trusted with, at least where AI only has to match error-prone humans, as in sales.
2 independent voices · 2 shows1 new this month
said Manny Medina (Topline), Amanda Kahlow (Topline, Revenue Builders)
5 sources
- Short, structured AI hackathons are an effective way to get teams building with AI.
2 independent voices · 2 shows
said Saumyo Mukherjee ([Un]Churned), Sam Jacobs (Topline)
3 sources
- AI tools substantially shorten software implementation cycles for customers.
2 independent voices · 2 shows1 new this month
said John Gilbo (Impact Pricing), Jason Goldsmith ([Un]Churned)
2 sources
- Data security and compliance concerns are a major brake on AI adoption and AI purchases at larger and regulated companies.
2 independent voices · 2 shows
said Sam Costello (Revenue Builders), Jeremey Donovan (The Revenue Leadership Podcast)
3 sources
- AI vendors need deep domain expertise built in, not technology expertise alone, to win.
2 independent voices · 2 shows
said Ghazi Masood ([Un]Churned), John Kaplan (Revenue Builders)
2 sources
- AI will shift managers and reps from reviewing data to acting on exception alerts and pre-prioritized escalations.
2 independent voices · 2 shows1 new this month
said Grant Clarke ([Un]Churned), Ian Tickle (The Revenue Leadership Podcast)
2 sources
- AI has collapsed the cost of producing and distributing marketing content.
2 independent voices · 2 shows1 new this month
said Ross Simmonds (The Dave Gerhardt Show), Aviv Canaani (The Revenue Leadership Podcast)
2 sources
- AI makes it easy to prototype most of a solution, but getting it production-ready takes most of the effort.
2 independent voices · 2 shows
said Simon Farthing ([Un]Churned), Alex Bilmes (Revenue Builders)
2 sources
- Incumbent platforms that hold the system of record will absorb AI rather than be displaced by AI-native entrants.
2 independent voices · 2 shows
said Kyle Norton (The Revenue Leadership Podcast), Manish Chawla ([Un]Churned)
3 sources
- AI work should start from a specific, named task with a defined output rather than a vague mandate to use more AI.
2 independent voices · 2 shows
said Jordan Crawford (Topline), John Kaplan (Revenue Builders)
4 sources
- Voice AI agents will take over routine phone interactions such as reservations and call-center information collection.
2 independent voices · 2 shows
said Glenn Fogel (Grit), Sahir Azam (Revenue Builders)
2 sources
- High-quality AI-made content still demands substantial human taste, selection and effort.
2 independent voices · 2 shows1 new this month
said Alex Mashrabov (The Twenty Minute VC), Marc Ferrentino (Topline)
2 sources
- Switching between model providers is easy, so the model layer has little customer stickiness.
2 independent voices · 2 shows1 new this month
said Alex Mashrabov (The Twenty Minute VC), AJ Bruno (Topline), Asad Zaman (Topline)
3 sources
- AI can flag churn risk much earlier by combining product usage and customer interaction signals.
4 independent voices · 1 show1 new this month
said Rebecca Nerad ([Un]Churned), Mark Vovsi ([Un]Churned), Adnan Rahman ([Un]Churned), Rob Edmondson ([Un]Churned)
5 sources
- Many AI-native companies' hypergrowth is inflated by experimental demand that will fade.
2 independent voices · 1 show
said Mark Roberge (Grit), Joubin Mirzadegan (Grit)
2 sources
- Current models already hold far more value than businesses use, so AI would transform industries even if progress stopped.
3 independent voices · 1 show
said Peter Zaffino (Grit), Bret Taylor (Grit), Arvind Jain (Grit)
4 sources
- AI context should be scoped to one job per thread or folder, because mixed or overlong context degrades output.
2 independent voices · 1 show
said Kyle Norton (The Revenue Leadership Podcast), Jordan Crawford (The Revenue Leadership Podcast)
3 sources
- AI will drive large productivity gains within the next one to two years, such as three- to five-fold gains in software engineering.
2 independent voices · 1 show
said Kyle Norton (Topline), AJ Bruno (Topline)
2 sources
- AI lets one person cover more domains, making customer-facing roles more generalist.
2 independent voices · 1 show
said Josh Schachter ([Un]Churned), Abbas Haider Ali ([Un]Churned)
4 sources
- Any AI platform you buy must let you see and extract the intelligence it builds, since walled-garden intelligence is a trap.
2 independent voices · 1 show
said Jonathan Moss (The Revenue Leadership Podcast), Kyle Norton (The Revenue Leadership Podcast)
2 sources
- AI customer agents will first serve smaller, long-tail customers that human teams cannot cover economically.
2 independent voices · 1 show
said Simon Farthing ([Un]Churned), Adnan Rahman ([Un]Churned)
2 sources
Actions written 10 Oct 2026 from the most useful of 636 recent insights and checked against them.
What was said 61 insights matching
Dev Ittycheria expects model safety, measured by some benchmark-like mechanism, to become a major enterprise buying criterion alongside cost. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Jason Lemkin estimates enterprise demand for tokens is probably an order of magnitude more than companies have figured out how to surface. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Dev Ittycheria expects Jevons Paradox to expand enterprise AI usage as costs fall, making the model market non-zero-sum. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Rory O'Driscoll predicts CFOs will force CIOs to cut token bills by routing non-frontier work to cheaper models. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Enterprises don't need frontier-level intelligence for every workload. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Enterprise executives at Dreamforce did not want to use Chinese open-weight models and did so only under cost pressure. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
A near-frontier US open-weight model that is 3–4x cheaper lets enterprises cut cost without using Chinese models. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Token cost and budgets plus IP and data rights are important factors in which AI models enterprises choose. Listen
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?”
Listen to the episode Episode AI Link to this Report a problem
One participant argued that until enterprises learn to sandbox end-to-end agentic redesigns of their operations, AI will diffuse at the usual enterprise pace of seven to eight years. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Manny Medina sees software companies that sell transformation alongside the product as the biggest unlock for enterprise AI adoption. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Enterprise AI spend is shifting from experimental budgets to longer commits, while overall penetration is still around 1%. Listen
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”
Listen to the episode Episode AI Link to this Report a problem
Lochmiller expects a mix of custom company models and closed frontier models, with neither approach dominating. Listen
Asked whether enterprises training their own models on private data will eat into frontier labs' business, he calls it possible but says there will always be demand for the frontier. Private data can be added to closed frontier models or to self-owned open-source models. He points to Cognition post-training its own model and Harvey possibly doing similar in legal, and notes the frontier labs are also pursuing domain-specific knowledge.
“I don't think it's going to be like one approach is going to be the dominant approach.”
Listen to the episode Episode AI Link to this Report a problem
Jack Altman expects more US enterprises to post-train their own models on open weights and run them through inference companies. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Asad Zaman expects gross margin pressure on AI application companies to persist, because buyers in banking, private equity and law will demand the best possible ROI and so force vendors onto frontier models. Listen
Asad Zaman said the application company's job is to build product on top of the best intelligence so the customer gets ROI. He does not believe customers at investment banks, private equity firms and law firms will accept less than the best potential ROI. Vendors will therefore be forced to keep paying for top models, leaving ongoing gross margin concerns that he called 'very concerning'.
“I don't think these customers that sit in investment banks and private equity firms and law firms are going to be looking to not have the very best potential ROI, which means you're going to find yourself forced to use these models to serve your customer.”
Listen to the episode Episode AI Link to this Report a problem
Moving from frontier models to self-trained open-weight models is a real downgrade in quality, however the move is marketed. Listen
Asad Zaman responded to Harvey's plan to train and host its own open-weight models by saying companies use fancy language about training to disguise the downgrade. He likened it to a firm that only hired from Harvard switching to a mid-tier university and claiming nothing changed. He added that law firms are 'trying to do deterministic work with probabilistic technology' and questioned how happy they would be to lose access to the best models.
“it's like going from hiring people only from Harvard to then going and hiring people from the worst university you can find or some mid -tier university and saying, It's the same thing. It's not the same thing.”
Listen to the episode Episode AI Link to this Report a problem
An AI tool will not replace enterprise sellers on $500,000-plus deals, because these require coordinating many stakeholders, though small deals can be fully automated. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
AI committees and AI addenda in buying processes can slip deals that looked committed, so reps need to track how each buyer's process has changed Listen
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”
Listen to the episode Episode AI Link to this Report a problem
AI helps with smaller, high-volume deals, but strategic and larger deals still need in-person selling, which Bove sees as a competitive edge. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Mark Vovsi gave an example of a chief customer officer asking about a customer's status without logging into the CSP. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Proofpoint's early warning system, which Mark Vovsi said was about to be released in its first version, combines structured, unstructured, human-reported, calculated and machine learning risk signals. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Cybersecurity companies must use AI now, because attackers face no governance limits and are already using open source models. Listen
He said attackers have no governance committees, data loss prevention rules or boards to report to, so they will use these tools however they can for ransomware and breaches. He said some attacks leveraging open source AI models have been reported in recent weeks. He said companies should understand how an adversary would use these models and should work with vendors that take AI seriously.
“They have no governance committees.”
Listen to the episode Episode AI Link to this Report a problem
Tenable's agentic AI layer is cutting some customers' remediation and prioritization work from weeks and months to minutes and hours. Listen
Mark said Tenable's agentic AI technology, Hexa, runs on top of Tenable One to simplify and modernize workflows, including mundane and repetitive manual tasks. He said tasks that once took a month or a month and a half can now be completed much faster by agents. He cited customers doing remediation and prioritization in minutes and hours instead of weeks and months.
“We have customers now doing some certain things around remediation and around prioritization in minutes and hours compared to weeks and months.”
Listen to the episode Episode AI Link to this Report a problem
Buyers arrive informed by AI, so sellers who skip AI-assisted research will be at a significant disadvantage. Listen
Mark said CISOs and CIOs do the same AI-assisted research on vendors, so they know the issues and the questions before meetings. He said that research is now easy, so reps who skip it are at a significant disadvantage. He was responding to John McMahon's point that AI now lets even lazy reps bring an informed point of view quickly.
“if you're not doing that legwork and homework, which is pretty damn easy, you are going to be at a significant disadvantage.”
Listen to the episode Episode AI Link to this Report a problem
Reps who use AI as a co-pilot free 15 to 25 percent of their week for customer time, against Tenable's targets of seven or eight face-to-face calls and 10 to 15 video calls. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Tenable's sales team gets its biggest AI gains on the back end, in sales operations, forecasting analysis, enablement and pre-call planning. Listen
Mark said the biggest leverage so far is on the back end, including getting quotes done in sales operations and analysing historical forecasting data and trends. He said enablement now records customer sessions and uses AI to tweak content in near real time. He said pre-call planning, which once meant reviewing many sources, now produces AI briefings that reps find resonate.
“where we are getting the biggest leverage out of it is really kind of on that back end.”
Listen to the episode Episode AI Link to this Report a problem
CISOs and CIOs now open many meetings by asking about a vendor's partnerships with frontier AI labs, before getting to the product. Listen
Mark said nearly every meeting with CISOs and CIOs starts with questions about partnering with Anthropic and OpenAI, and then moves to how AI changes the threat landscape. He said boards are asking these buyers to prepare for AI-enabled attacks, and that buyers also raise models such as Glasswing and Mythos. He said it is hard to discuss a specific part of a platform without addressing that context first.
“Even if you're going in for a meeting, talking about a specific, you know, part of your platform, they want to start off with, hey, what's it like partnering with Anthropic?”
Listen to the episode Episode AI Link to this Report a problem
Agarwal predicts the next wave of AI value comes from entirely new work rather than doing existing work faster or cheaper. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Tim Rutten predicts banks will eventually run open-source models on-premises to contain risk and control unit costs. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Intelligence is not the bottleneck for banking and that banks need smaller, more efficient models with better infrastructure. Listen
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.”
Listen to the episode Episode AI Link to this Report a problem
Tim Rutten cites Lloyds and JPMorgan as examples of measured AI returns, with Lloyds at about 50 million last year and tracking to 100 or 200 million this year. Listen
He said Lloyds Banking Group reported about 50 million in AI ROI last year and is tracking to 100 million, if not 200 million, by the end of this year. The transcript then notes this comes from maybe 0.1% of the bank's total operation, meaning adoption is still early. He also said JPMorgan Chase released financial statements showing serious returns from its programs.
“tracking to 100, if not 200 by end of this year”
Listen to the episode Episode AI Link to this Report a problem