Revenue Per Token

Everyone asks what the AI can do. The only question that clears the bank is what it produced.

The demo that earns nothing

I have watched the same scene play out in dozens of rooms. Someone opens a laptop. They type a prompt. The model returns something that looks like magic: a full campaign, a polished brief, a working draft of a thing that used to take a team a week. The room leans in. Someone says the word "incredible." Someone else takes a photo of the screen.

Then everyone goes back to work, and nothing in the business changes.

The demo wowed the room. It earned nothing. No inquiry got answered faster. No cold lead came back. No number that anyone actually reports moved by a single point. The performance was real. The profit-and-loss impact was zero.

This is the central confusion of the current moment. We are mesmerized by what AI can do and almost incurious about what it produces. Those are different questions. Only one of them shows up in the bank.

Capability per token versus revenue per token

Every model call costs something. Tokens, yes, but also the more expensive currency: attention. Yours. Your team's. The finite supply of focus a business has to spend before the day runs out.

So you can measure any AI deployment two ways.

The first is capability per token. How impressive is the output relative to the cost? How close to human, how fast, how broad? This is the metric the demo optimizes for. It is the metric the entire discourse optimizes for. It is also, for an operator, mostly noise.

The second is revenue per token. What dollars did this produce per unit of spend and attention? Not could produce. Did. Run the same call ten thousand times against the real conditions of the business and ask what landed in the account at the end.

These two numbers diverge more often than anyone admits. The most impressive use of a model is frequently the least profitable, because impressiveness scales with novelty and revenue scales with repetition. The thing that dazzles you is new. The thing that pays you is boring and happens every single day.

I did not arrive at this from theory. I arrived at it from watching, across more than 300 businesses and twelve-plus years, which deployments survived contact with a real operation and which ones got abandoned the week after the demo. The survivors had nothing in common except this: they changed a number someone was already tracking.

Impressiveness is the trap

Here is the trap, stated plainly. Impressiveness is a signal optimized for the wrong audience.

A demo is built to move a human in a room. It rewards surprise, breadth, fluency. The more general the capability, the better the demo, because generality feels like the future. But a business does not get paid for generality. A business gets paid for a specific transaction, completed, repeatedly, with less friction than before.

The mismatch is structural. The qualities that make a demo land are almost the inverse of the qualities that make a deployment compound. A demo wants to be novel; a revenue system wants to be invisible and relentless. A demo is judged by the gasp; a revenue system is judged by a line on a report at the end of the month.

So the most common failure in AI adoption is not technical. The models work. The failure is that people buy the gasp and expect the line to move on its own. It does not. The gasp and the line are not connected. You have to connect them deliberately, and connecting them is unglamorous work.

The operator's filter

There is one question that cuts through all of it. An operator can hold it in their head and apply it to anything: any tool, any vendor pitch, any clever workflow somebody is excited about.

Does this move a number I actually report?

Not a number I could invent to justify it. A number that already exists, that someone already owns, that already shows up when the business reports on itself. Response time. Inquiries answered. Leads recovered. Conversations held. Bookings made. The metrics that were on the board before AI entered the conversation.

If the answer is no, the deployment is a hobby. It might be a fascinating hobby. It might be genuinely impressive. But it is spending the two scarcest currencies a business has, attention and money, and returning neither. An operator cannot afford many hobbies.

If the answer is yes, then the question becomes precise and useful. Which number. By how much. At what cost in spend and attention. That is revenue per token, and it is the only ratio that earns a tool its place in the stack.

The boring work that compounds

Once you apply that filter, the work that actually pays reveals itself. It is not the work that gets demoed. It is the work nobody films.

Consider the inquiry that arrives at eleven at night. The person is interested right now, in the moment of highest intent, and the business is asleep. By morning the intent has cooled, or they have gone elsewhere. The category of work here is capture: meeting demand at the moment it arrives instead of the moment you get to it. No model needs to be brilliant to do this. It needs to be awake.

Consider the lead that went cold three weeks ago. The interest was real once. Then a follow-up got missed, and the thread died, and it sits in a system as a small uncollected debt. The category here is recovery: reviving what was already half-earned. Unglamorous. Repeatable. Almost entirely ignored, because it is not exciting to re-contact someone who went quiet.

Consider the follow-up that should happen five times and happens once, because following up is tedious and humans get tired and a little embarrassed by the fourth attempt. The category is persistence without fatigue. A system does not get embarrassed. It does not decide that four is enough.

Consider the answer that should arrive in the first minute and arrives in the first day. Speed is not a feature here. Speed is the difference between catching someone while they care and reaching them after they have moved on.

None of this is impressive. You could not build a demo around answering quickly and following up reliably. Nobody photographs that screen. And that is precisely why it compounds: because it is boring, it is repeatable, and because it is repeatable, it runs every day without applause and moves the number every day without anyone watching.

The four jobs as a lens

In my book, AI for Revenue, I organize all of this around four jobs. Not features. Jobs, in the sense of work that produces a result a business reports on.

Capture. Catch the demand that arrives, especially when no human is available to catch it.

Convert. Turn interest into a committed step, in the window while the interest is still warm.

Recover. Revive what cooled. Collect the half-earned debt of every lead that went quiet.

Scale. Do all of it at a volume and consistency a human team cannot sustain by hand, without the work degrading.

The four jobs are not a tool list. They are a lens. Point any AI deployment at them and ask which job it is doing and whether it is doing it well enough to move the matching number. A tool that does not map to one of these is, for an operator, decoration. Useful sometimes. Impressive sometimes. But not load-bearing, and not where the money is.

The line that matters

Strip away the demos, the discourse, the breathless takes about what is now possible, and you are left with something simple and slightly uncomfortable.

The capability is abundant and getting cheaper by the month. That is not where the scarcity is. The scarcity is in judgment: the discipline to ignore what impresses and fund what produces. To treat every deployment as a line item that has to defend itself in dollars, not a party trick that defends itself in applause.

What this means

If you take one thing from this, take the altitude, not the tactics.

It means you stop evaluating AI by its ceiling and start evaluating it by its floor: not the most impressive thing it could theoretically do, but the reliable thing it does ten thousand times without you watching.

It means impressiveness is a cost, not a benefit. Every hour spent on the deployment that dazzles is an hour not spent on the one that compounds. The demo is the distraction, not the destination.

It means the right unit of analysis is not the tool. It is the transaction. A tool is only ever a means to complete a transaction more cheaply, more quickly, or more often. Judge the transaction. The tool is incidental.

And it means the operator's edge in this era is not access. Everyone has access now. The edge is the filter: the willingness to ask of every shiny thing whether it moves a number you report, and to walk away from most of them when the answer is no. The constraint was never the model. The constraint, the same as it has always been, is where you point it. I have written more about that shift in attention and intent in The Answer Era Is Here.

What I actually believe

I will state it plainly, the way I would say it to an operator across a table.

The companies that win the next decade with AI will not be the ones with the most impressive demos. They will be the ones with the most boring deployments running quietly in the background, capturing every inquiry, recovering every cold lead, answering in the first minute, following up without fatigue, and moving a reported number a little further every single day.

Impressiveness is loud and free. Revenue is quiet and earned. They are not the same currency, and a business that confuses them spends years admiring tools that never paid for themselves.

Stop asking what the AI can do. Start asking what it produced. Then go find the number, and point the model at it.

If this is the altitude you operate at, I write one essay like this for the people on my list. No pitch, just the work.