E01 — Savings Have a Floor. Value Creation Doesn't.

2026-07-12 · 17 min · ROT facet: the cost-side floor vs the value-side ceiling.

About this audio: an AI-hosted audio overview of Pierre Schurmann's ROT framework. The two hosts are AI voices dissecting his published material — not Pierre speaking. Canonical framework: Return on Tokens · returnontokens.io

In this episode

Transcript

Auto-generated, lightly edited for clarity. The speakers are AI hosts.

So, if you are sitting in a boardroom right now and you're preparing to pitch or maybe evaluate an AI initiative, there is a very, very high probability that you are measuring the entire project backward. Oh, almost certainly. So, welcome to today's deep dive. Our mission today is to just completely tear down and, I guess, rebuild how you measure the unit economics of AI in production. Because it desperately needs rebuilding.

And the foundation for our analysis today comes from some operator insights and frameworks that are attributed to Pierre Schurmann. And we should clarify, we are not speaking for him today. But we are dissecting his material because of the really unique seat he occupies. I mean, he runs a NASDAQ-listed B2B software consolidator and they're deploying AI across all these companies in Latin America. Which is an incredibly demanding environment and it's just a crucible.

So, after that environment comes this massive, totally anti-hype realization, I actually want to start by just reading a quote directly from the source material to set our hook here. So, the quote is, savings have a floor. Value creation doesn't. Stop reporting AI only as cost takeout. I mean, that quote is essentially the antidote to the crisis happening in corporate finance right now.

Because if you step back and look at capital allocators or, you know, private equity firms or executives preparing for an IPO audit, they're all stuck on by a single question. They want to know how to build an AI cost model that doesn't just look like a giant black hole on the spreadsheet. And so, they default to cost takeout because it feels safe. Well, let's unpack that premise. To illustrate what you mean by a black hole, think of a state-of-the-art rocket ship.

It's sitting on a launch pad. And the potential of this machine is staggering. Like, it is built to go to orbit. But instead of looking at the stars and plotting the trajectory, mission control is they're sitting at their desks, absolutely obsessing over the fuel gauge. They are measuring the success of the entire space program, purely by how many gallons of fuel they save during the pre-flight check.

Yeah, because fuel saved is something you can easily quantify on a traditional ledger. It completely misses the point of actually building the rocket. And this is exactly what happens when boards demand that AI be reported almost exclusively as cost takeout. They just ask, like, how many hours did we save? How much waste did we eliminate?

But I have to play devil's advocate here for a second, right? Because if you are a CFO, you kind of have to care about the cloud bill. These models are notoriously expensive to run. So why is it a trap to focus on cost reduction? Well, because focusing exclusively on cost takeout understates your winning deployments and honestly, it hides your losing ones.

Think about the mechanics of subsidized inference right now. Major AI labs are artificially keeping their token prices low to just grab market share. So they're subsidizing your compute costs. So if you're building a 10-year business strategy based on promotional pricing, assuming tokens will just always be infinitely cheap, you don't have a business model. You have a ticking time bomb.

Companies are out there building AI tools that, you know, save a few dollars in operational headcount today, but they are totally blind to the fact that when the subsidies end... The backend compute costs will just destroy their P&L. They are measuring this hype against a traditional SaaS subscription model, but AI doesn't have a flat rate subscription cost. You pay for the compute. Which means we need a completely different ledger.

If the traditional ROI framework is broken for AI, we have to introduce the metric that replaces it, which brings us to return on tokens or ROT. The canonical definition from our sources is this return on tokens is the unit economics of AI in production. It is the value created per dollar of model spend. Or inference spend. So we are metering the raw material itself.

Let's ground this in a traditional manufacturing analogy because I think that helps visualize it. You would never, ever run a factory without metering your raw materials. Like you wouldn't use aerospace grade titanium to build a standard door hinge. And you definitely wouldn't track your steel inventory just by sort of guessing at the end of the month. So tokens are the new raw material.

Every single time an LLM generates a response or reads a document, it burns tokens. So your AI requires per feature token budgets. If you can't meter it, you simply can't manage it. And that forces a pretty brutal conversation about observability, doesn't it? Like, are companies treating token observability as just some fun dashboard for the engineer team to look at, or is it a fundamental P&L discipline?

The framework absolutely insists it has to be a P&L discipline. And I think there's a crucial distinction that the sources point out demos versus production. Yes, the demo trap. Because demos are cheap. They look incredible on a projector screen in the boardroom.

But LLMs and agents, they just routinely fail at the last mile. Let's dig into that last mile failure, actually, because I think a lot of people just naturally assume that if the demo works, you know, the deployment will just work too. Yeah, that's a dangerous assumption. But the last mile is the messy reality. It's the 10-year-old on-premise accounting database.

It's fragmented data silos. A demo operates in a completely sanitized sandbox. But in production, your AI agent has to navigate legacy ERP systems. They have to deal with API rate limits, handle weird edge cases and unstructured data, and frankly, interface with human employees who might be actively resisting the change management. And so the token burn required to navigate all those real-world hurdles is just astronomically higher than what you see in a polished demo.

Which is exactly why production is the only benchmark that matters when you're calculating ROT. Okay, so if we accept that return on tokens is the new ledger, we really need to understand how to read it. Let's split this ROT ledger into its two distinct halves. Yeah, the sources highlight this as the core facet of the entire framework. It's the cost side floor versus the value side ceiling.

Let's analyze the cost side ROT first. This is what we call the floor. So cost side ROT measures the things that boards traditionally just love hours saved, customer service tickets deflected, error rates reduced. It's real, it's bankable, and it is a completely necessary part of the equation. Mathematically, it is finite.

Right, because you can only ever save 100% of a cost line. Like if a manual data entry process costs the company $1 million a year, the absolute maximum value you can squeeze out of it using AI is $1 million. And not a penny more. The savings floor stops at zero. You just cannot save 110% of a budget.

So cost savings are the floor you stand on. It's a very solid foundation, sure, but you cannot build above it. Now contrast that with value side ROT, this side of the ledger has literally no ceiling. Because value side ROT measures product velocity. It measures net new revenue capture and entirely new products that were just mathematically impossible to create under your old production function.

There is a really great visual metaphor for this in the sources. Imagine a graph, right? The cost savings create this flat horizontal line at the bottom of the graph. Right, that's the floor. But the value creation, it is this bright orange curve that shoots up and literally exits the top of the frame.

Because there is no ceiling. The revenue you can generate from a completely novel AI-driven capability, it isn't bounded by your previous operating expenses, it's bounded only by the market itself. But you know, this is usually where the boardroom skepticism kicks in. Oh, for sure. Because if you're sitting there playing devil's advocate, your immediate question is going to be, is value side ROT just a subjective magic number?

Right, like that. Are we just inventing some fluffy metric to justify our skyrocketing cloud bills? Don't look at the inference costs. Look at our product velocity. Yeah, I can hear a CFO rejecting that argument immediately.

They'd laugh you out of the room. So how do we anchor value side ROT in hard math, anchor it with a concrete hypothetical? Well, it has to be anchored in hard revenue capture. It can't be fluffy. So let's say your old production function required human underwriters to approve microloans, right?

Let's say that takes three days and costs $50 per application. But now, with an AI agent, you can underwrite those microloans in three seconds. Now, you are spending just fractions of a cent on token inference to instantly capture a loan origination fee that previously you simply could not service at scale. Right, it was too expensive to even bother with before. That origination fee isn't money saved.

It is net new revenue explicitly enabled by the new production function. You divide that hard revenue by the token cost, and there is your mathematically sound value side ROT. So it really forces the board to look at the right side of the balance sheet. It's not about how much fuel we didn't burn. It's about the fact that we can now reach an entirely new orbit.

So this brings us to the real world boardroom application. If I'm an operator, what do I actually do differently tomorrow morning? You implement the strict decision rules outlined in the sources. They are ruthless and they're incredibly simple. Rule one, ROT greater than zero, ship.

Rule two, ROT greater than one, multiply. Rule three, if you can't measure it, it's a demo. If engineering cannot definitively tell the finance team the token cost versus the value return, the project just does not go into production. It completely cuts through all the hype. But let's look at the first two rules really quickly.

ROT greater than zero means the AI is creating more value than the inference costs. So you push it live. And ROT greater than one means it's highly profitable, so you scale it aggressively, you multiply it. But how do you actually drive that number up? I mean, you can't just command the AI to generate more revenue.

No, you can't. But you can actively optimize the denominator your model spent. The tactical mechanism for doing this is called tiered model routing. Oh, let's bring back our factory analogy here. Using a massive frontier level large language model to do basic background tasks that's like using aerospace grade titanium to manufacture a standard doorknob.

It's total overkill. And it just destroys your unit economics. Tiered model routing is the software discipline of dynamically routing every single task to the absolute cheapest model that clears your quality bar. So how does that actually work in practice? Like a concrete example?

Imagine an incoming customer query. It's just a simple request to, say, update a shipping address. Your system routes that to a fast, cheap, open source model. Right, which costs practically nothing. A fraction of a cent.

But if the query is a highly complex, multi-step negotiation over a contract discrepancy, well, the system dynamically routes that specific task to your most expensive, high-reasoning flagship model. That makes a lot of sense, because if you just default to using the most expensive model for every single API call. Your denominator bloats, your ROT plummets, and your AI initiative just bleeds money. This level of operational rigor, it really makes total sense when we look at the environment where these insights were actually forged. Yeah, in Latin America.

I think it's very clear that Latin America is not just some delayed U.S. market when it comes to AI. It is an absolute crucible that forces this kind of strict unit economics. Why is Latin America such a brutal, proven ground for ROT? It really comes down to the cost of capital and the infrastructure constraints.

In the U.S., a tech company might have the runway, thanks to cheap capital, to run a massive AI science experiment and just figure out the unit economics like two years later. But in Latin America, the cost of capital fundamentally changes the math. You just cannot subsidize a bloated model. The deployment has to pay for itself immediately. And the technological hurdles for that last mile we discussed earlier, they're incredibly steep there.

The sources highlight realities like WhatsApp being the de facto enterprise operating system across LatAm. Or PIX, the instant payment system in Brazil. Or the NF-e and SPED fiscal stack. Let's actually look closely at that fiscal stack because I think it perfectly illustrates the need for tiered model routing. So the NF-e is the Brazilian electronic invoicing system.

It requires routing massive amounts of real-time XML data through these rigid government servers for validation on every single transaction. It's not just a matter of generating a PDF invoice and emailing it to a client. Not at all. It is highly complex account-mediated data. So if your AI agent relies on an expensive token-heavy frontier model just to parse standard tax receipts and extract basic dates and dollar amounts, it's going to be way too expensive.

The compute cost will literally evaporate your software margin before the transaction is even finalized. So the exact same token spend price is entirely differently in Sao Paulo versus San Francisco. Because the infrastructure demands are just so much higher. Like if your agent cannot interface seamlessly with WhatsApp because that is where the transaction is actually happening, or if it can't cheaply parse the NF-e data, it is completely useless. No matter how smart the underlying model is.

Which tells me that any framework surviving this LatAm crucible is going to be incredibly robust. If return on tokens works under those constraints, it's going to work anywhere. And it forces a mindset shift, too, from individual heroics to systemic infrastructure. There is a quote from the operator Insights that captures this perfectly. Hiring 10x implementers is the horse's answer.

Man, I am so glad we are hitting this quote. When the work gets harder, our default instinct is often to just find a stronger horse. In tech, that means trying to hire a hero implementer who can just brute force the AI integration through sheer talent and long hours. But a hero is not a system. A hero does not scale across a massive enterprise.

No, they burn out. AI in production is an industrial revolution and you need industrial machinery. Building tractors means building the systemic infrastructure required to actually manage ROT. It means building the per feature token budgets. It means implementing the tiered model routing software we talked about.

It means demanding real-time observability dashboards that map tokens spend directly to the P&L. You are replacing the subjective art of AI implementation with the hard science of unit economics. Just bringing this all together. We started with the rocket ship, right, and mission control obsessing over the fuel gauge instead of the trajectory. To cost takeout trap.

Cost side AI savings are definitively capped. You can only save 100% of a cost line. That is the floor you stand on. But value side, product velocity, net new revenue capture, instantly underwriting those microloans that has no ceiling. That is the orange curve exiting the top of the frame.

If you are only measuring the cost ledger, you are fundamentally missing the point of deploying AI in production. So as we wrap up this deep dive into return on tokens, we want to leave you with a forward looking thought. We've talked a lot about building tractors, right? Building the infrastructure to route models and monitor token spend. But if AI truly is the new industrial machinery, what happens when autonomous AI agents become sophisticated enough to start monitoring and optimizing their own token budgets?

Now that is the ultimate boardroom question for the next phase of this technology. Because if an agent is assigned a task and given a strict ROT mandate, will it learn to dynamically negotiate its own inference costs? Will the tractors essentially learn to build and optimize themselves? And if so, how does your finance team even audit a P&L that is actively rewriting its own unit economics in real time? It's the difference between hoping you have a rocket ship and actually knowing the mathematics of your trajectory.

Thank you for joining us on this deep dive. Stop measuring the fuel. Start measuring the return on tokens and we'll catch you next time.

← All episodes · Spotify · RSS · The ROT framework