Fabien Cros on Pricing the Agent Economy
Fabien Cros is Chief Data and AI Officer at Ducker Carlisle and founder of SparkWise Data & AI (now Ducker Carlisle Data & AI). He was formerly Data & AI Country Lead for Manufacturing at Google. He spoke with Steven Forth for the upcoming book "Pricing for the Agent Economy," co-authored with Michael Mansard and Wolfgang Laga. His remarks below have been lightly edited for clarity and put into the first person.
Pull Quote
"The traditional value-based model — doing it the right way, really quantifying the value and creating the differentiation — will be very relevant, and in a sense it will be easier, because you take away the psychology."
Key Takeaways
Value-based pricing finally becomes real. For decades, "value-based pricing" was really the illusion of perceived value, because it was priced for humans. Machines don't care about human psychology — they respond to quantified value — so pricing has to get mathematically rigorous in a way it never had to before.
Discounts and promotions lose their power. Promotions and discounts work because they exploit human psychology ("I'm giving you a deal so you'll buy from me"). Machines don't respond to that framing — they just compare prices. What looks like a discount becomes simply a transaction price, or there's no deal at all.
Credits force real cost discipline. Building a credit system means translating an intangible (tokens, compute, wisdom, intelligence) into a tangible unit a buyer can budget against. That forces companies — the way it forced Google — to get rigorous about their true cost structure, not just their price.
Electricity, not compute, is the real bottleneck. As intelligence becomes cheap and abundant, the scarce resource shifts to energy. Whoever controls electricity capacity — Cros points to China's investment there — will have the durable advantage, and future pricing and consumption may ultimately be measured in energy terms.
The hardest problem isn't technical, it's societal. Automating tasks and deploying agents is comparatively easy. The genuinely hard problem is fairness and equity: avoiding a world where technology concentrates power and resources in very few hands while leaving much of the population without a role.
My Story
I've spent quite some time in the AI space. I started my career at Accenture in the US — I worked there for seven years. Toward the end of my time there, I was spending a lot of my time on AI, and my leadership eventually told me it was a waste of my time, that we'd never sell projects around AI. So I ran away and joined a startup called Artefact, based in Paris, specialized in data and AI. We scaled that business really well and sold it to a private equity fund called Ardian, which sold it again five years later for $1.1 billion.
Unfortunately for me, I exited after the first sale. I went to Google, where I led Data and AI for manufacturing for a couple of years, and then decided to go back to entrepreneurship. I started a company with a bunch of friends from Artefact and Google. We got acquired and financed by a US private equity fund, and now we're part of the Ducker Carlisle group, where I lead the data and AI practice as Chief AI Officer.
We're a 200-person consulting company specialized in industrial manufacturing, automotive, equipment, and building and construction. That's really our niche, and within it we do M&A advisory, supply chain consulting, strategy consulting, pricing research, a lot of market research — and, through my team, data and AI services.
It's very relevant to what this book is about, because we're building the agents that are creating this agent economy. Take a manufacturing company — it's very rare that they have a strong AI team internally. So they call us and explain what kind of agent they need, or at least the idea they have, and our job is to build the product.
We leverage a lot of the Google DNA and mindset — we build products the Google way, with real user centricity, as software developers leveraging concepts from DevOps. But our specialty is really AI.
What "Agent" Means to Me
Everybody argues over the definition. For my clients, I try to keep it simple. There's the automation piece — workflows, automated decisions — that's something that started in the '50s and has gotten better and better. Now, with LLMs, automation is easier because we can ingest information and the output can be different from the input.
The "agent" part is autonomy — when a machine can take a decision and an action. It's not the same input always producing the same output; you give something to the machine, and the machine can choose to send an email, make a phone call, prepare a quote — really create something that has an impact on the outside world.
I'm sure some people will agree with that framing and some will disagree, but that's how I try to explain it.
Technically, you can build agents inside Copilot or inside Claude. I don't have the exact numbers, but I'd be surprised if there weren't more agents inside those ecosystems than outside them.
I think the excitement is that this is the first time we can give a machine a task with a real degree of autonomy — and that gets into the difference between generative AI and discriminative AI. This community of AI has been around a long time — machine learning, deep learning — but the input and output were similar in kind. In classic machine learning, you give numbers and get numbers back. Even with a learning component, it's numbers to numbers.
Now you can send a couple of words and get back a picture or a video. That's the beauty of the agent: this capability to ingest an input, understand the input and the intent, and create an output that matches that intent — even when the output is a completely different type or nature than the input. In a way, that's how humans work. We're a big machine with a lot of sensors, mixing feelings, vision, and words as input, and the output could be a feeling, a word, or a movement.
So I think we're getting very close to mirroring how humans work. That said, I'm a big fan of Yann LeCun here, because I don't want to give the impression that this is AGI. LLMs aren't intelligent — they're very good probabilistic machines trained on the internet. We've nailed text input and output, and found a way to translate pictures into text, and now we're doing the same with text and video. But it's still grounded in the probabilistic calculation of the next word, which is very different from a machine — or a human — based on causality.
A human isn't calculating probabilities to produce an output. A human processes data and creates causality: if I keep getting punched in the face by something, I'll move away from it. Honestly, if the training set told an LLM "you get punched in the face and you don't move," the machine would say exactly that — you get punched and you don't move. There's no concept of feeling or understanding. It's a purely probabilistic calculation.
How I Approach Pricing Agents
Yes — I was literally with Michael yesterday talking about exactly that.
I'm stealing with pride here — I think it was Steven and Michael who designed this "impossible triangle" with the cost vertex. I think it's really good, because it explains why you can't continue to price an AI agent like SaaS, or by seat — the value lever is completely different; it's a different value metric. I use that a lot to explain that we need to change.
The one limitation I flagged with Michael is that we don't control the cost of tokens. In this field, it's actually fairly easy for me to get predictability on my own costs — I can move away from frontier models, go open source, run my own server, or buy my own hardware, and then I get something closer to linear cost, which is different from software (which is close to flat) but also different from the exponential cost risk you can get with frontier models. So there's still a risk that this cost vertex becomes irrelevant as things progress — I don't know, because I'm genuinely puzzled here; we'll see how it plays out. The assumption is that labs will keep competing with each other and that open source will always be available. That's a big assumption — especially with how good the Chinese labs are getting. I think it'll continue, but who knows.
Given that assumption, though, I think the real difference will come from the true application of value-based pricing. We've spent decades talking about "value-based," but when you really think about it, it wasn't truly value-based — it was the illusion of the perception of value, because we were selling to humans. Tell a human "you'll get the information in nanoseconds instead of milliseconds" and they don't care — they can't perceive the difference. Tell a machine that, and it's a huge impact. That's why the explanation of value, and the edge you get from being genuinely value-based in your pricing, becomes a real game-changer — change one component and you can completely erase your market share, because you're dealing with a machine.
We're moving away from relationship-based selling. In some ways, I think the blueprint already exists — look at advertising, and specifically what Google does with its bidding process. That's the only truly value-based mechanism I've seen at scale, because it's a purely math-based mechanism with a buyer, a seller, and a transaction price fully determined by a set of criteria and inputs every single time. In the past, only Google could afford to build that, because of its massive scale, and that scale was what made it a trusted intermediary between buyer and seller. I think most companies will be forced into that same space, because as we rely more on agents, we'll need our agents to transact directly with each other — and because things are moving so fast, we won't be able to stay the bottleneck. We'll be obliged to give our agents a way to transact.
That's what makes this exciting from a pricing perspective: how do you redesign pricing models to let your agent explain what you want and what matters to you, and negotiate that between two machines? I think the traditional value-based model — done the right way, with real quantification of value and real differentiation — becomes very relevant, and in some ways easier, because you take the psychology out of it.
Now systems, solutions, agents, whatever we want to call them, need to become much more transparent about value and price. And if you follow that line further, it gets really exciting, because concepts like discounts and promotions will start to go away. As a machine, if you use promotions, the machine will just game you and wait for the next promotion — so it kills the impact of promotions entirely. Same with discounts, which are really based on human perception: "I'm giving you a discount so you'll buy from me." A machine doesn't care whether you're trying to get it to buy from you or not. So it's either you lower your price — and it's not a discount, it's simply a transaction price — or there's no deal. I think that concept of transaction pricing is going to be very interesting.
What Credits Mean to Me
For me, it goes back to how we priced things at Google. From a Google Cloud point of view, it's really the concept of consumption plus commitment — commitment was a big thing — and when you commit, you get credits: an amount of something you can consume, whether that's tokens, storage space, compute power, or electricity.
That's interesting, because today I'd say it's really only the most advanced tech companies using credits, because it's complex — you need a mechanism to translate your output into a credit and explain that translation, which takes real heavy lifting, thinking, and analysis. At Google, you could do that because you had hundreds of people doing pricing. But if you're selling something simpler — say a pizza — I'm not sure you can break the pizza down into credits the same way. But at the end of the day, when you buy a pizza, you're buying a sum of credits that adds up to a pizza. So from a machine's point of view, it's actually pretty easy to break down that price — maybe the credit is calories. You're paying $15 for, say, 3,000 calories.
Yes, the data aspect was for sure part of it. The compute piece, I'm less sure about — we priced storage with credits too, not just compute.
But what were the implications of credit-based pricing for Google, and how might that impact other companies? For me, it's almost like an electricity mindset — they said, "We're selling something very intangible, so how do we track this intangible?" They had to go back to first principles, and I think the only way was to explain that one credit gets you XYZ. That makes it very tangible for an IT department: "I need Y units of that, which means I need X credits" — and now you can budget. You can't really buy servers in that way, because you need to scale up and down, and you can't price by server because sometimes you're sharing one. The unit of measurement is genuinely hard to pin down.
That's very similar to AI. What's the unit of measure? Some of my clients are selling wisdom, not intelligence, as a consultant — so how do you measure wisdom? You try to quantify it: maybe it's a sum of years of experience, so my team has, say, 30 years of experience, and if we can quantify the time spent, maybe that becomes the credit. But it's mentally hard to translate, because we're human — we're not designed to think in a single unit of measurement. We see a pizza, not a sum of atoms, even though from an atomic point of view a pizza really is just a sum of atoms. We price what we see.
With something like an LLM, where you don't "see" the output in the same tangible way... I think what could be genuinely interesting is that we'll finally get very good at understanding costs. Trying to think in credits, as a way to price the output of a world where intelligence itself is nearly free, forces you to go back to first principles about what things actually cost. Intelligence is cheap, but you still need materials, electricity, and computing power — so credit-based thinking forces much more diligence on cost breakdown. You can't go to a client and say, "here are the 50 ingredients you need" — you package it so it's one credit and they understand it, because that's what they're looking for. But internally, as a company, you need that full breakdown. That rigor requirement could be pretty revolutionary.
Where This Is Going
We have the illusion that things are moving fast, but honestly, when I look at adoption relative to the technology's real potential, I think we're still massively under-served. We still haven't automated a lot of things that could be automated — everyone talks about robots, but we haven't really seen anything truly impactful yet. So I don't think five years out will be radically different; we'll make good progress and be much better equipped. But — and this is more a hope, maybe more a dream than a prediction — I think we'll separate two things.
We'll separate the illusion of unlimited access to workforce and intelligence from the real bottleneck, which I believe will be electricity. At the end of the day, it comes down to the cost of electricity, and I think China understands that much better than Europe or the US right now — they have far more electricity capacity than we do, so for that reason alone they may well win. We'll probably end up measuring things in terms of electricity need, or computing power — but given the progress we're making in computing power itself, I think electricity will be the real bottleneck. Even with nuclear power, we don't have unlimited energy. So eventually we'll be able to quantify things much like we quantify a car's fuel consumption: my agent is consuming X, and the output is Y.
It'll become very easy to compare efficiency, power, and intelligence based on input and output. And at the same time, I think the one truly non-renewable resource for humans will be the cost of time. There's a movie called In Time where people are no longer paid in money, they're paid in time. That's the trade that's coming: if you have unlimited access to labor and intelligence, the real question becomes where you choose to spend your time. Time becomes almost a currency in the way the dollar has been a currency — a way of trading time for monetary value.
Advice for the Next Generation
I'll be upfront that my answer is biased by where I spend my time. For me, it's that we're still running companies on 19th-century concepts. All the incentive structures, all the frameworks we use, are becoming irrelevant — and we're this strange, almost schizophrenic generation: we have access to the technology, but we're still measured on things that no longer make sense. In board meetings, I'm always the one asking, "how do you decide who gets promoted to the board?" And the answer is usually, "whoever has the biggest team" — and I think, that makes no sense anymore.
When you look at incentive structures, do we accept a 10x difference in incentives? Usually the answer is no, we need to stay within a certain band. But we're moving into a world where you'll have 100 people and one person doing 99% of the value — yet you still need the other 99, because the old 80/20 rule just re-applies itself to whatever's left even after you cut the bottom 80%. I don't think we're ready to handle that. So my honest answer is that people entering this field need to invest a lot of time in building a new mental framework for thinking about the technological impact on society — this idea of the top 1%, this question of how you price something that has no human interaction at all, how you connect and filter information from your own "cyber workforce" to someone else's cyber workforce.
Unfortunately, I think the technology is taking over the world in a specific sense: in the past you had data people, pricing people, and finance people as separate functions. Now you have no choice but to be a finance-and-data person, a pricing-and-data person, a marketing-and-data person. That's going to be genuinely hard, because we're asking functions that were never interested in data to become data-fluent. Finance-and-data will be relatively easy. But how do you create creatives who are data-driven? How do you create journalists who are data-driven? How do you find writers who think that way? It's going to be very hard.
And honestly, I'm quite pessimistic from a societal point of view, because we're building technology that, by design, concentrates power and resources into very few hands. I'm almost tempted to tell young people: don't learn to code, go study philosophy instead — because I think the biggest problem we're going to have to solve isn't technical at all. It's fairness, equity — how do we live together with another species, which is the machine? How do you build a world that's fair for humans and fair for everyone? For me, that's the hardest problem to solve. Automating factory work, putting robots on the floor — those are almost the easy problems by comparison. The hard problem is making sure you don't create a huge population that becomes useless to society. I think that's where bad things happen — or where wars happen.
Pricing for the Planet
With Stephan Liozu, I co-founded Pricing for the Planet — a kind of think tank and education platform for exploring how we should use pricing for sustainability, and how we build pricing practices that are genuinely sustainable. I think this connects directly to the credit conversation. If we connect philosophy and sustainability, there's a high chance that we'll need to find real money to repair the damage we've done to nature — and if we decouple taxation from the labor workforce, that money will most likely have to come from this digital, cyber economy.
It wouldn't surprise me if one or several governments eventually agree to tax every transaction between machines, with the proceeds going into a fund — hopefully for sustainability, or at least as a kind of VAT-at-scale for government revenue. And coming back to an earlier point, the useful side effect is that you're forced to genuinely understand your cost structure, which means you can genuinely understand your impact on nature. Right now there's often no data, so nobody can really tell you the true cost of something — all the externalities included. But as we put agents throughout the value chain — everything we're building and extracting — we can get full visibility: how many trees did it take to build an iPhone, or for a hotel stay? You can trace it all the way back: every time you extract from a forest, you should be paying to regenerate that soil.
Actually, thinking about credits again, I suspect credit-based systems will reopen the conversation about blockchain, because you probably need a global ledger — unbiased, and owned and shared across the whole economy — to track credits properly. That starts to look like a kind of token-based organization. We're heading toward a currency conversation, which honestly could be its own book.
I absolutely think credits could become fungible and transferable between organizations — especially if we get to a level where there's a clear correlation between what I'm selling and its equivalent in energy consumption. At that point, it becomes a currency in its own right — you don't need dollars anymore, because the value is backed by real energy consumption.
This transcript has been condensed and lightly edited for clarity and flow, and adapted into a first-person narrative. Interview conducted July 24, 2026 for "Pricing for the Agent Economy" by Steven Forth, Michael Mansard, and Wolfgang Ulaga.