Topline · 4 Oct 2026 · From the week of 28 September
Turning AI Tokens Into ROI (That You Can Measure) | CEO @ Paid, Manny Medina
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Manny Medina, co-founder and CEO of Paid (a monetization platform for AI agents) and former founder/CEO of Outreach, joins Topline hosts Sam Jacobs, AJ Bruno and Asad Zaman in person. They discuss how agents should be priced once seats stop making sense, why tokens are the wrong billing unit, and a value ladder that starts with an itemized, explainable bill and only later reaches ROI. Manny argues that 'how do I price my agent' is really an NRR problem: fixed-price deals win logos easily but stall at the first renewal. He also covers lessons from Outreach (product marketing, raising a round he didn't need, culture versus winning), Paid's consultant-led, take-rate go-to-market, and his prediction that software gross margins will settle around 40% to 60% rather than SaaS's 70% to 80%.
For founders
- Manny predicts all software will be agentic within five years, and says SaaS incumbents are best placed to build agents on their own data and workflows; if they don't capture that value, someone else's agents running on their MCP layer will.
- Manny says fixed-price agent deals make new logos easy but leave early-stage companies boxed in by Series A or B, when the first renewal arrives and there is no room to expand.
- Manny says he regrets raising Outreach's last round because the company didn't need the money, and argues founders should always shop an inbound term sheet rather than respond to exploding deadlines.
- Manny calls SaaS's 70% to 80% gross margins an anomaly of cheap money and expects normal software to run on 40% gross margins, 60% at most, with narrow vertical agents able to be highly profitable in their niches.
- Manny's operating principles at Paid are a small high-performing team, revenue growing faster than headcount, operating leverage early and high ACV, and he regrets not hiring a product marketer early this time.
For revenue leaders
- Manny argues that charging for agent work rather than seats removes the renewal contraction where procurement cuts 100 seats to 50 because only 20 are heavily used.
- Manny says agent builders must make the bill explainable before arguing ROI: itemize the research, decks or analyses produced in plain terms, the way an old phone bill listed every call, rather than billing in tokens.
- Manny says most agent pricing today sits between effort and output, because true outcome pricing needs both autonomy and attribution, and attribution is hard outside areas like support, procurement and accounts payable.
- Manny says outcome definitions differ per customer, so vendors need a catalog of outcomes and flexible metering; he cites Sierra as pricing on outcomes from the start with a catalog of outcomes.
- Paid qualifies inbound leads by asking whether pricing is a CEO-level problem, opens with a monetization consultant before selling software, and charges a percentage of the customer's revenue for the life of the contract.
What was said 38, most useful first
Pricing agent work instead of seats removes the renewal contraction that happens when only a fraction of purchased seats are heavily used. Listen
In SaaS you sell 100 seats, procurement comes back at renewal saying only 20 are heavily used, and the customer buys 50, killing the add-on you planned to sell and threatening an RFP if you push back. With agents you charge for all the work done. If only two of 100 users are hyper-users, the vendor doesn't care, because those users drive value for the whole organization and the vendor is paid for that value.
“if you've sold 100 seats and only two were hyper-users, you don't care because you're charging for all the uses and all the value those two seats were getting”
Manny Medina describes a value ladder for agent billing: transparency first, then explaining the value of each piece of work, then customer-specific ROI. Listen
Step one is letting the buyer and their finance team see what the agent did, described in normal terms, like the old itemized phone bill where you went straight to the biggest line to check it. Step two is explaining the value, for example that a piece of research would have taken a human a long time, or that the agent produced a PowerPoint or an analysis. ROI comes last and depends on each customer's own equation. He says agents he works with are often failing at the first rung, and that once work is visible a buyer can question how artifacts were used, but cannot dispute that the work was done.
“So explainability is like the first rung in the ladder.”
True outcome-based pricing depends on autonomy and attribution, and because attribution is hard, most agents today are priced between effort and output. Listen
Manny frames two boundary conditions: whether the agent did the work without a human (autonomy) and whether the work can be attributed entirely to the agent (attribution). He says outcome pricing is visible in support, inventory management, procurement and accounts payable, where the agent completes the transaction itself. Most other use cases lack real attribution, so pricing sits between effort (a little above token consumption) and output (a produced artifact such as a presentation or email), or a combination of both.
“Attribution is really hard. So this is why true outcome-based pricing is going to take some time.”
The definition of an outcome changes for every customer, so agent metering has to be configurable per customer. Listen
Manny's examples for a support resolution: the agent answered and the customer went away; the customer went away and didn't return within a week, which means keeping a window open to watch for the signal; or the case was escalated but a human responded within two minutes. He says this flexibility in understanding what the agent did became Paid's differentiator.
“So every single definition of an outcome or work changes per customer. You need a flexible platform to do that.”
Paid moved from cost-and-margin observability to agent metering after finding that the buyers who cared about margins were not fast-growing companies. Listen
Paid initially looked more like Datadog than a metering platform, answering both what work was done and what it cost so customers could protect margins. The customers interested in margins were private equity-style businesses, which Manny says were not the right customers. Paid kept the cost-management piece but shifted its focus to flexible metering of agent work.
“I love you guys, but it's not the right type of customer. They're not in the game right now. I want companies that are growing fast.”
'how do I price my agent' really means 'how do I grow net revenue retention faster', because new logos are easy and expansion is hard. Listen
Enterprises are currently buying one of each agent, so new-logo acquisition is easy, and many companies took fixed-price deals to grab that money. The problem shows up when they go for expansion revenue on a fixed price. Manny says a sustainable business has to flex with the value it delivers, and a vendor capped by fixed-price contracts won't capture that value; someone else will.
“You know, the enterprises are buying one of each, so new logo is not hard. Expansion is hard.”
Paid opens sales engagements with a monetization consultant who sketches the customer's pricing design before Paid sells any software. Listen
Manny made himself a promise after Outreach to solve the customer's whole problem rather than sell tools. Paid brings in a consultant with a background at firms like Simon-Kucher or McKinsey who does a quick back-of-the-envelope pricing design, then co-designs and implements, a forward-deployed model he compares to Harvey's. He won't sell the consulting on its own to customers of other metering platforms; it comes only with the product.
“I'm going to bring a monetization consultant. Somebody has worked at Simon-Kucher, McKinsey, BCG, or whatever, and he's going to sit down with you and do a quick back of the envelope design of how your pricing should be before I sell you anything.”
Manny Medina chose to build Paid as a system of record rather than a workflow tool, because workflow products face constant churn risk. Listen
Paid owns the record of what the agent did, the credits charged and the auditability of both, which Manny calls a financial record. He acknowledges a higher bar, since billing errors and credit misassignments are a real risk when code is increasingly writing itself. He says workflow tools are like being a shark: always moving, with customers able to churn at any minute.
“I don't want to build workflows anymore because workflows is like being a shark.”
Early-stage agent companies selling fixed price don't feel a pricing problem until their first renewal, usually around Series A or B, so Paid doesn't target them before then. Listen
Early-stage companies take logos quickly at fixed prices and box themselves in. When the first renewal conversation comes, often around Series A or B, they realize the deals signed the year before are about to renew with no room to grow. Paid doesn't engage companies that haven't felt this pain yet.
“So pricing is not a problem if you're just taking down the logos at a fixed price.”
Listen to the episode Pipeline & demand generation Link to this
Manny Medina predicts normal software businesses will run on about 40% gross margins, 60% at most, calling SaaS's 70% to 80% an anomaly of cheap money. Listen
At Amazon, which he joined after HBS, Manny had to memorize operating margins: books 20%, electronics 7%, jewelry 3%. He cites Bezos's line 'your margin is my opportunity'. Coming from that, he found Microsoft's and SaaS's 70% to 80% margins a luxury that cheap money made possible and that wasn't even spent efficiently on growth. He still says gross margins must be positive.
“I feel like a normal, you know, software business will be running on 40% margins, 60% tops.”
Agent pricing should be customized to each customer's definition of value, citing Sierra's catalog of outcomes as the model. Listen
Manny says Sierra went to outcome-based pricing from the start and, rather than debating what counts as an outcome, keeps a catalog of outcomes already tallied for other customers; he says their margins are great as a result. He rejects a host's suggestion that this only works in support. He describes the belief that pricing must fit in one standard box as a limiting belief.
“you should make your pricing way more customizable, depending on what your customer value is.”
Manny Medina predicts that no software will be non-agentic within five years, and that SaaS incumbents, not upstarts, are the primary beneficiaries. Listen
Manny says he used to argue around San Francisco that SaaS was dead and would take the shift lying down, but has changed his mind. SaaS companies already have knowledge and workflows embedded in their software, which makes them best placed to build agents on top of their stack. Their challenge is figuring out how to capture value, and if they don't, someone else will.
“So they are the primary beneficiaries of building agents on top of their stack. And they just need to figure out how to capture value. And if they don't, somebody else will.”
MCP layers on SaaS products recreate the Salesforce API ecosystem dynamic, forcing every SaaS company to choose between capturing value from MCP calls and building its own agents. Listen
Manny compares the moment to when Salesforce opened its APIs and an ecosystem of go-to-market tools, including Outreach, was built on top. Every SaaS company now has an MCP layer that any agent can run on. He says this race will force everyone to build some kind of agentic workflow on top of their offering and work out how to make money from it, because it is not the same as charging per seat.
“every single SaaS company has now an MCP layer on top of it, which makes it available to any agent to run on top of that. Now, the question is, how do you capture value for that?”
SaaS buyers never seriously asked what value seats delivered, and that the onus is now on agent builders to answer that question. Listen
Manny argues that seats felt so intuitive that nobody asked what 100 seats actually bought, and vendors answered with usage charts, which he dismisses. Buyers are now asking value questions of agents, which he considers proper. He says agent builders have to show buyers what they get for the agent.
“And the first chart that everyone wanted to show up is usage. And who gives a fuck about usage? Like you want value.”
Buyers' concerns about budget predictability under usage-based agent pricing can be handled with caps, limits, wallets and assignments. Listen
When a host raised CFO concerns about managing budgets and forecasting spend under consumption pricing, Manny called it a totally fair question. He listed caps, limits, wallets and assignments as the mechanisms. He treats this as the second question after showing value.
“There's caps. There is limits. There is wallets. There's assignments.”
Tokens should never be the unit agent builders charge on, because they are an internal measure that means nothing to the buyer. Listen
Manny says tokens vary by lab and are an internal measure of how LLMs are built. A host noted that two models can do the same task using different numbers of tokens, which Manny agreed with. He says agent builders, unlike the labs, have to show a measure of the value delivered to the customer.
“A token should not be a unit of value for charging because a token is really kind of like the inside secret.”
Manny Medina repeats an analogy he heard that frontier labs behave like hardware companies, releasing expensive models and then cutting cost so distilled models never catch up, but he is unsure this holds as compute costs rise. Listen
In this view, labs release the most expensive model, then make it more token-efficient and lower the token price. Distilled models take six to nine months to catch up and never do, because the frontier model is already cheaper by then and a newer one is out. Manny cautioned that higher interest rates and rising GPU-hour and compute costs could change this, and said he doesn't know how it plays out.
“They're really hardware companies that their job is to put out the most expensive model out there and then use that knowledge to bring the cost down over time and then do it again and again and again.”
A competitor that moves to cheaper fine-tuned open-source models to protect margins hands rivals a strong counter-position: always using the best model. Listen
A host argued that companies fine-tuning open-source models through inference providers will be a year to a year and a half behind the frontier, leaving room for a competitor built on the best model. Manny, quoting a previous Topline episode, framed it as two Harveys: one optimizes margins, and the other promises never to put your case law into the cheap model. Another host pushed back that buyers often use cheaper lawyers for routine work.
“I am using the best model and I will never put your case law into the cheap model.”
Asad Zaman predicts that buyers will start asking service providers which AI model they use, and that the answer will become a differentiator. Listen
Asad compares it to hiring a driver for the windy roads to resorts in northern Pakistan, where the first question is which car they'll use. He thinks that as companies adopt AI, customers will ask what model or product a provider uses and will be concerned if it's a weak one.
“there's going to be a new question of what are you using like what product. using and that's going to become a differentiator”
Paid charges customers a take rate on their revenue for the life of the contract instead of a software fee. Listen
Manny says customers have been open to the model. A host called it the cleanest form of outcome-based pricing, and Manny summed it up as 'you grow, I grow'. He later said a take-rate model is one reason he isn't worried about investor questions on growth.
“We take a percentage of revenue, so we always take a take rate.”
He should not have raised Outreach's last round, which he took under market pressure as a first-time CEO rather than out of need. Listen
Manny says the company didn't need the money and he got carried along by the market and the momentum. He now takes a more deliberate approach to fundraising at Paid, noting that previous rounds determine the next one. A host added that Outreach's strong narrative made it easy to get caught up in that hype cycle.
“I felt like that I shouldn't have raised the last round. And the reason was we didn't need the money.”
Manny Medina advises founders never to accept an inbound term sheet without shopping it, and to ignore exploding deadlines. Listen
An inbound term sheet signals that other investors would also offer one. Manny cites Bending Spoons saying their acquisition offers stay valid at the same price a month later, and argues a VC's valuation shouldn't change in a week either. He also says he prefers big-brand investors as partners.
“you should never take an inbound term sheet because if somebody send you an inbound term sheet, that means that there's other people out there that will give you a term sheet.”
A company can choose when its fast-growth story starts, so building slower foundations first is not a handicap. Listen
Asked how investors react to Paid selling into enterprise and building a system of record, both slower than the AI companies going from zero to $100M in months, Manny says he doesn't worry about it. Take-rate companies are among the fastest growing, and nobody discusses where a fast-growth curve starts: you build, then lock in and go fast. He cites NVIDIA as a fast-growth story that took 30 years to make.
“So you can make that starting point anytime you want.”
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”
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.”
A 3% to 5% agent error rate on multi-step tasks is a troubleshooting problem, not a reason to limit what agents are trusted with. Listen
A host argued that even frontier models fail on 3% to 5% of multi-step tasks, which limits what enterprises can trust AI to do, even with forward-deployed engineers on top. Manny replied that teams had the same problem before and that AI isn't a magic bullet. Frequent misfires usually point to a harness, memory or context problem that you debug until it's fixed. He says the goal is an organization that needs fewer people, has more throughput or delivers higher quality.
“when you're seeing a lot of errors or a lot of misfires, then you may have a harness problem, maybe you have a memory problem, maybe you have a context problem”
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”
One participant reported that a New York biotech investor says his job now takes 25% of the time it used to, using only Claude and ChatGPT. Listen
The investor wasn't using specialized tools, just basic workflows while switching between Claude and ChatGPT. He now spends 25% of his former time on the same work, freeing 75% for other things. The speaker presented it as an example of how well AI works where it fits.
“He's like, my job is now 25% of what it used to be.”
Paid qualifies its mostly inbound pipeline by asking whether pricing is a CEO-level problem, and walks away if it isn't. Listen
Most of Paid's demand is inbound from companies struggling to price agents. If the problem isn't owned by the CEO, Paid punts; if it is, Manny gets on the call with the CEO. He says these conversations are usually triggered by board discussions. He doesn't try to predict which customers will win.
“So our question is more like, you know, is this a CEO problem or not? If it's not a CEO problem, we punt.”
One participant said the blended gross margin figure they hear most for AI companies by year-end is 25% to 30%. Listen
The speaker linked this to models improving fundamentally every year, which can produce stretches of negative margin followed by price increases and rationalization. The result is a blended yearly gross margin around 25% to 30%. They gave it as a figure they hear from others, not their own data.
“The number that everybody I hear from is that at the end of the year, 30, 25 to 30”
Manny Medina expects many hyper-verticalized agent companies to be modest in size but very profitable, like Veeva in pharma. Listen
Manny sees agent companies solving very narrow problems in very narrow industries. He argues that owning the whole niche makes them very profitable even if they aren't huge. He cites Veeva, which does a few things only for pharmaceuticals and turned out to be a huge business.
“So they're solving very narrow problems in very narrow industries.”
Manny Medina is seeing some agent application companies retreat to selling just the harness because customers want to bring their own LLM. Listen
Paid's customers usually build an agent and run it. Lately, Manny says, some have moved toward selling only the harness, because buyers say they want to use the product with their own LLM. He flagged this as a separate topic and didn't develop it further.
“Because a customer will be like, fine, I want to use you, but I want to use my LLM. So they will be like, you know, sell me the harness.”
Manny Medina regrets not hiring a product marketer early at Paid, having had one early at Outreach. Listen
At Paid, Manny decided he could handle positioning and enablement himself. He now feels he is working all the time because of it. At Outreach, a product marketer was one of his first hires and he says his life was 'glorious'. He describes this as a wrong lesson he took from reflecting on the past.
“I should have done that because I did that at Outreach and my life was glorious.”
Manny Medina is running Paid on the principles he wanted at Outreach: a small high-performing team, revenue growing faster than headcount, early operating leverage and high ACV. Listen
Manny lists the goals he set for this company and says he is achieving them. He wanted a small, high-performing team, revenue growing faster than headcount, operating leverage early, and high ACV because 'it's an easier life to live'. He also wanted more time with customers. He adds that being a second-time founder in this platform shift has huge advantages.
“I wanted to keep revenue growing higher than growing headcount. I wanted to hit operating leverage early. I wanted to keep ACV high because it's easier life to live.”
Manny Medina recounts Satya Nadella telling him that life is path dependent, which he uses to set aside regret over Outreach decisions. Listen
As Microsoft's M12 fund's highest-performing company, Outreach earned Manny a keynote slot with Satya Nadella. In the green room Manny spent ten minutes on what had gone wrong and what he should have done. Nadella replied that he was only there because of the things he had done. Manny says that if he hadn't made those choices he wouldn't be building Paid, which he loves.
“Manny life is path dependent You're only here because of the things that you did if you didn't do those things you wouldn't be here.”
His Outreach lesson is that culture serves winning, and that a company is better framed as a sports team than a family. Listen
Manny says that at Outreach culture sometimes came ahead of winning, and winning should have been the main thing. A CEO's biggest job is still to bring everyone along and keep them aligned, but culture shouldn't be confused with winning. Culture feels much better when you're winning.
“But at the end of the day, it's a sports analogy more than a family analogy. We're all here to perform.”
Manny Medina agrees that reimagining work for AI means starting from the outcome, since every intermediate process step is negotiable. Listen
A host asked whether the clearest form of reimagining is to sell the outcome rather than the form-filling work, citing how factories were redesigned around electricity. Manny agreed. Another host added that chat interfaces should give way to software that already knows what to do.
“At the end of the day, the process exists for a reason. And that reason is the outcome. And all the middle steps are all negotiable.”
Manny Medina sees Paid's market, like Outreach's early market, as an evangelism problem: convincing buyers that a solution exists. Listen
Outreach had to convince the world that sales could be a workflow: a pipeline that can be generated, where enough coverage means you make the number. Manny says that wasn't clear back then and became clear over time. At Paid the job is to evangelize that agent pricing can be solved, with work that is observable and auditable. He notes the constraint is scale, since travel to CEO events takes time.
“It was convincing the world that sales could be a workflow”