Ten Years of Outcome-Based Pricing: Lessons from Afiniti
Interview conducted by Steven Forth, July 27, 2026
Pull Quotes
"Coming to your clients and saying 'I'm going to sell you this fantastic AI model' doesn't truly mean anything if you don't have a business impact. It doesn't mean anything if you're not able to prove it. And it doesn't mean anything if you don't believe enough in what you're selling to say, 'Look, I'm just going to take a share of what I deliver for you.'"
"The most rational value proposition (you only pay for what you get, this is money on the table, just take it) doesn't always work inside organizations. Some clients will simply prefer to pay a price tag without knowing exactly what they're going to get."
"If you don't get perfect attribution, it's harder to justify your value share. Attribution is the core of the issue when you think about outcome-based pricing."
"The counterfactual is the product. Everything else is a story about the product."
Key Points
Pay-for-performance only works when the value is measured and proven inside the customer's own systems. At Afiniti, on/off sampling, running the AI pairing engine for twenty minutes, then reverting to the incumbent first-in-first-out logic for five, created a continuous, auditable benchmark of the lift the system delivered, and Afiniti was paid only on that lift.
The counterfactual is the expensive part, and most agent vendors are skipping it. Afiniti ran a permanent production holdout for the life of the contract, switching its own engine off one minute in five to keep the benchmark honest. Without that discipline, outcome-based pricing for agents is usage pricing under another name.
Attribution is the core of the issue in outcome-based pricing. Translating a business outcome into measurable fields in a customer's systems takes enormous time and effort, and even then, outcomes like 30-day churn are multi-factor and hard to attribute to a single interaction. Imperfect attribution weakens the value-share argument, and can even corrupt the data an AI model learns from.
Credits are the survivable form of outcome pricing. Afiniti rarely sold the pure model; floors, caps, budgets and industrial pilots made it workable for both sides, and credits generalize that instinct. The open design question is how much outcome alignment you can layer on a consumption model before the cost of measuring it exceeds what it buys you.
Under outcome pricing, the vendor pays to fix problems it did not cause. Performance dips at Afiniti usually traced back to client-side data issues or unannounced production changes, and the vendor absorbed the diagnosis and the remediation while earning less that month. Floors and caps protect the vendor as much as they reassure the buyer.
Outcome pricing is as hard to implement inside the vendor as inside the customer. When revenue lands well after signature and its size is unknown on the day you sign, no commission plan is clean: pay on signature and you pay for pipeline, pay on outcomes and your best people leave before the money arrives.
Budget predictability is an underestimated obstacle to outcome-based pricing. When there is no price tag, there is no budget line, no project code, and no allocated stakeholders, paradoxically, "free" can mean nobody cares. Credits, floors, caps, and industrial-scale pilots are practical mechanisms for restoring predictability.
The most rational value proposition does not always win. Even "you only pay for what you get" meets resistance inside large enterprises. With no single buyer, seller, or budget, success depends on executive sponsorship and an internal champion who sells the value to other stakeholders.
AI's biggest opportunity may be in the "boring economy". Applying an AI-transformation mindset to conventional, non-tech businesses can unlock efficiency and entirely new customer offers, and greater capability makes human responsibility more important, not less.
In Nicolas Gallot's Words
From investment banking to an AI startup
I started in investment banking after graduating from HEC Paris, working at Morgan Stanley. I spent three years there and then moved on to private equity. And then one day a former colleague called me and said, "Look, I have this crazy company. Someone is offering me the chance to build the French subsidiary, and I think we should do it together."
That was ten years ago, and the company was Afiniti. I knew nothing about artificial intelligence and nothing about contact centers, which is the world this company operates in. What I could read was the business model, and it was unusually smart: entirely based on pay for performance. That was enough for me. If a company is confident enough to be paid only on the value it creates, the value is probably there, and someone with an analytical background can help prove it, sell it and scale it. So I jumped in.
We built the French subsidiary from nothing, then I took on responsibilities at the international level, and I stayed through the recapitalization and change of ownership at the end of 2024 and the transition that followed, until I left at the end of 2025.
Ten years building and running a business around managed AI services is what I bring to this book: a practitioner's perspective. My experience is narrow, one use case in one industry, and I am genuinely looking forward to reading the theory. But I can share what actually happens when you sell pay-for-performance and outcome-based pricing for a decade, across many countries.
As for what I am doing now: I moved on in early 2026 to start a venture I have wanted to do for some time, acquisition entrepreneurship in France. I am looking for an established mid-sized French company to acquire, run and develop, almost certainly with no connection to AI or contact centers. In a sense I am doing again exactly what I did ten years ago: entering an industry I do not know, on the strength of what I can read in a business model, and betting that I can add something. The difference is that this time I will own the outcome.
The Afiniti model: getting paid only on the lift
Let's start with the basics. When you are an AI company, you have value to offer your clients, and your models are based on data. Coming to your clients and saying "I'm going to sell you this fantastic AI model" doesn't truly mean anything if you don't have a business impact. It doesn't mean anything if you're not able to prove it. And it doesn't mean anything if you don't believe enough in what you're selling to say, "Look, I'm just going to take a share of what I deliver for you." That was the story Afiniti was built on and the one we told for ten years, and it was genuinely innovative, especially a decade ago.
Let me explain what we actually did, because it helps make this concrete. What we were doing was quite simple, in the contact center world. Think about your telecom, banking, or insurance provider. You call to change your plan, add a new one, or maybe to move to a competitor. In the world without Afiniti, your call would be routed to a group of representatives trained to handle your request, but the exact person you talked to depended entirely on the time of your call. It was strictly first in, first out. Call five seconds earlier or five seconds later, and you would talk to a different person.
What Afiniti set out to solve was exactly that. Large enterprises invest so much to improve outcomes on contact center calls, and there is enormous value in retaining a customer or adding a plan, plans that stay in place for 30, 50, 100 months afterward. The lifetime value is huge. So our value proposition was: we can find the best pairing between the representative and the caller. Instead of first-in-first-out allocation, you get an allocation based on behavior, the best pair available in the queue, without anyone noticing anything, because the difference is typically plus or minus three or four seconds. To do this we used data such as the historical outcomes of each representative over their past 10,000 interactions, and on the caller side, the 100 to 300 records you typically have on each customer in the CRM.
There is a lot of data you can use to find the patterns of success, or of failure. And it comes down to a simple idea: human interactions matter. What Afiniti was really doing was finding those patterns of behavior and making sure we could form the best pairs to maximize the value of the calls.
That sounds beautiful, but Afiniti went further: we actually measured the impact of this one-to-one pairing, using on-and-off call samples. For twenty minutes, Afiniti would decide who talked to whom; for five minutes, we would revert to the incumbent FIFO logic. Twenty minutes on, five minutes off, all month long. At the end of the month you measure, for example, the conversion rate "on" versus the conversion rate "off," and you can see the lift. And the last mile was this: Afiniti offered to be paid only on that lift, measured and proven inside the customer's own systems.
People sometimes ask why we sliced by time rather than randomizing call by call. The answer is operational. To get a clean counterfactual you have to restore the incumbent state of the entire queue, not just of one call, because representative availability is shared across every caller waiting. Time slicing introduces an artefact of its own, since call mix shifts through the day, which is why the on and off windows had to alternate continuously all month long rather than being grouped into blocks. That level of detail is where these models are won or lost.
The issues nobody anticipates
That sounds like the perfect outcome-based model, and it sounds easy to sell. But I can tell you from experience that it opens up many issues people don't anticipate when they think about outcome-based pricing.
The first is budget predictability, which, interestingly, is something the industry is now also trying to solve with credits. Inside organizations, when there is no budget and no clear price tag on a project, nobody actually cares. It's "free." If there is no skin in the game, who should care about it? This is especially true for cross-functional projects: you don't get someone from the IT team allocated, because there is no budget, no project code name inside the company. That kind of inefficiency was one of the key lessons for me.
The second took me longer to accept. Having operated in this field for ten years, I learned that in large enterprises, the most rational value proposition, "you only pay for what you get; this is money on the table, just take it", doesn't always work, for many different reasons. Even when you position your offer with outcome-based pricing and think you now have the most compelling value proposition possible, you will find that some clients think differently. Sometimes they simply prefer to pay a price tag, without knowing exactly what they will get, rather than being assured up front that they will only pay for what they receive.
The third issue sits inside the vendor rather than the customer, and it is the one that surprised me most. The vendor carries the downside, and that is the whole point of the model. What is less obvious is where the downside actually comes from. When our performance dipped, the cause was rarely the model itself. It was usually a data problem, or a change the client had made in production without anyone telling us. So we would put experts on it, spend weeks diagnosing, and fix a problem we had not created, in a month where we were earning less precisely because of it. Under outcome pricing the vendor ends up being the party who cares most about the customer's data quality and change management, which is an odd form of alignment. It works, and it is expensive. Floors and caps were as much about protecting us from that exposure as about giving the client a predictable budget.
The fourth is the one nobody talks about. Outcome-based pricing does something brutal to your own sales organization. If revenue on a deal only starts landing well after signature, and its size is genuinely unknown on the day you sign, how do you compensate the person who signed it? Pay on signature and you are paying for pipeline. Pay on realized outcomes and your best people leave long before the money arrives. We iterated on this for years and never found a clean answer, only workable compromises. Anyone moving to outcome pricing should budget as much design effort for their own commission plan as for the customer-facing model. The internal problem is the same size as the external one, and it gets a fraction of the attention.
Tactics that worked: credits, caps, and champions
So how did we overcome those objections? This is where thinking about credits, and the credit consumption model, becomes interesting. It typically wasn't framed as "we want to spend X amount on this project next year." But it could mean adding a floor or a cap. It could mean starting with a pilot, a very industrial pilot, where we deploy everything, look at what it actually delivers in production, and then allocate a budget appropriate to the project.
One thing we offered was simply: put a number on it. Set a budget, because it helps down the line on the project. And if we reached a point where we risked exceeding that budget, we would flip the system to 80% off and 20% on, instead of 80% on and 20% off, while the client worked internally to secure more budget. That kind of mechanism is how we managed to convince clients.
You also realize that not every company is the same. What was critical, probably as in any enterprise sale, but even more so when you need to align so many stakeholders and there is no clear single buyer, single seller, and single budget, was finding the right sponsorship at the executive level, and the right champion who could push the project inside the organization, really selling the value internally to the other stakeholders. Because this is about selling value, not selling a label, not "a $30 million AI project" where we talk about what we spend but never about what we deliver. In our case, it was about the value we were going to create, with a share of that value going to Afiniti, as opposed to putting a price tag on top.
Why outcome-based pricing hasn't taken off: attribution
Looking at the broader business landscape today: over the past few months we saw a lot of excitement around outcome-based pricing, and then it settled back into a more balanced view. What I observe in practice is that very little of it has actually shipped. And one of the key reasons, one that generalizes well beyond the Afiniti lesson, is attribution, in two senses.
First, the time it takes. What we built at Afiniti was a performance measurement document, going really deep into the customer's systems, bridging business understanding to data details, translating a business outcome, and the understanding of that outcome, into measurable fields and processes inside the customer's systems. That takes time and a lot of effort for a client, and again you need to align many stakeholders.
Second, attribution is genuinely hard to read. Even with the most accurate possible performance measurement document and countless hours discussing marginal cases, when you measure something like 30-day churn post-call, whether the client churned within thirty days of the call, so many things can happen between the call and the moment they leave. Can you truly attribute the churn to one particular interaction? Attribution is hard because outcomes are multi-factor. And if you don't get perfect attribution, it's harder to justify your value share, because then your value is kind of your value, but not just your value. In our case that raised the stakes considerably. The off samples did double duty: they were the measurement baseline, and they were also the only source of un-optimized data the model could keep learning from. So measurement error does not simply cost you revenue in a given month, it degrades the training signal itself. Attribution quality and model quality turn out to be the same problem.
Within the scope of my experience I am confident about this much: attribution is the core of the issue when you think about outcome-based pricing. Every other obstacle is downstream of it.
What this means for agents
Here is the test I would apply to any vendor claiming outcome-based pricing for agents. The pairing engine was Afiniti's visible product. What made the pricing model possible was the willingness to switch that engine off for one minute in five, permanently, in production, for the life of the contract. That permanent holdout is what turned a claim into a measurement, and it is the expensive part. I have yet to see an agent vendor willing to do it. Turning your agent off twenty percent of the time is commercially painful, operationally awkward, and it exposes you to a number you may not like. Until vendors accept that discipline, most of what is being described as outcome-based pricing for agents is usage pricing wearing an outcome-shaped label. The counterfactual is the product. Everything else is a story about the product.
Steven asked me about credit-based pricing, and I think my answer is implied by everything above. In practice we did not sell outcome pricing in its pure form. What we sold was outcome pricing wrapped in the mechanisms that made it survivable: a floor, a cap, a number the client could write into a budget, an industrial pilot before any commitment. Those mechanisms exist because pure outcome pricing is uncomfortable for the buyer's budget process and exposes the vendor to costs it does not control. Credits are a more elegant version of the same instinct, and I would not be surprised if that is where the agent market settles. I would not read that as a retreat. The interesting design question is how much outcome alignment you can layer on top of a consumption model before the cost of measuring it exceeds what it buys you.
Looking ahead: responsibility, and the boring economy
On the longer horizon I disagree with one of the loudest claims. I do not believe AI takes everyone's jobs. What I do see is enormous potential for efficiency almost everywhere. And I see something else that gets much less attention. All of this capability does nothing to move responsibility. Someone still has to tell the model what to do, and someone still has to answer for what it did. As the future owner of a company, that is the part I think about most. The more powerful the tools, the more the accountability concentrates in the person holding them, and the fewer excuses that person has.
I am optimistic about where this goes, even though I find the pace hard to read. Every quarter I feel like I am operating differently from the one before. Even in acquisition entrepreneurship, which is about as far from frontier technology as you can get, my working methods keep changing under me. The clearest shift is that analysis has become close to free, and once that happens it stops being the scarce resource. What remains scarce is access to good businesses, and the judgment to know which parts of it matter. I have had to get comfortable with the fact that the work I was trained to do is the part that got cheap. That is a strange feeling for someone who came out of investment banking, and I think a lot of people are about to have it. The habit I hold onto is refusing to do something the way I did it last quarter simply because it worked last quarter. I have no clear visibility on where this ends, and I am comfortable with that.
And here is what reassures me. When I say I want to acquire a company, I am looking for one that software has simply not reached yet. No tech, no platform, no AI, at least as it stands today. I am convinced that bringing an AI-ready mindset into that kind of business opens far more room than applying it to companies already saturated with technology, because the baseline is so much lower. Efficiency first, then new ways of working, and eventually new offers to customers. I believe in the potential of transforming the boring economy, and it is one of the positions I am pushing in my search.