Every AI budget cap I've seen this quarter is attached to a person. Apple cut some engineers from three hundred dollars a day to two hundred. Uber capped individuals at fifteen hundred a month. The pattern is everywhere, and it misses the thing it's trying to control, because almost none of the spend that worries a CFO is attached to a person anymore. It's attached to a workflow. An agent that reconciles a week of transactions overnight isn't anyone's seat. A pipeline that drafts, reviews, and files a hundred contracts belongs to a process, not a login. Budget by headcount and you can't see that spend, can't govern it, and can't tell whether it was worth it. We call the habit the seat reflex. It's the instinct to budget AI per person because software was always sold per person, carried into an era where the spend follows the work.
What is the seat reflex?
It's budgeting AI by the person using it, inherited from decades of per-seat software, applied to a cost that now scales with workflows instead of users.
For thirty years the unit of software spend was the seat. You bought a license per employee, the cost scaled with headcount, and finance could forecast the whole thing from the hiring plan. That's the model every company's AI budget was built on, and it broke this year. Industry survey data shows enterprises planned to roughly double AI spend in 2026 and ended up spending nearly five times the prior year, about two and a half times what they budgeted. That gap isn't waste. It's what happens when you plan for seats and get billed for tokens. The consumption didn't come from more people. It came from agents doing more work, in parallel, after everyone went home, and a per-seat budget has no line for that.
Why does a per-person cap fail?
Because it governs the wrong thing. It limits what an individual can spend while leaving the real consumers, the workflows, unmeasured.
Picture the cap in practice. An engineer hits her daily limit at three in the afternoon and stops. That's the behavior the cap was designed to produce, and we've written about what it costs in the ambition cap, where a ceiling tells everyone to stop reaching. The seat reflex is a different failure, one level down. Even a cap set generously can't touch the spend that matters, because that spend doesn't run under anyone's login. The agent fleet runs under a service account. The pipeline runs under the team's key. The overnight job bills to a cost center nobody reviews. So the cap squeezes the one place spend is visible, individual experimentation, and leaves the place it's actually growing, production workflows, exactly as unmanaged as before. You get a tidier number on the seats and a surprise on the invoice.
Isn't this just the expense reflex again?
No. The expense reflex is about which column AI sits in. The seat reflex is about which unit it's counted in.
We've argued that AI belongs in the capital column, not the expense column, in intelligence as capital. You can win that argument and still have a budget that can't see where the capital went, because the ledger counts people and the capital flows through workflows. The two reflexes compound. A company managing AI as an expense will cut it. A company managing AI per seat can't tell what to cut, so it cuts the seats, which is the part that was never the problem. Fix the column and you'll allocate deliberately. Fix the unit and you'll know what you allocated to.
What's the right unit?
The workflow, because that's where the outcome is, and cost only means anything next to an outcome.
The number a leader actually wants is what it costs to produce a finished result, the measure we've called cost per outcome. A reconciled ledger. A filed contract. A resolved ticket. Those outcomes belong to workflows, and so should the budget. When spend is tagged to the workflow that generated it, three things become possible that per-seat budgeting can't do. You can see which workflows are expensive relative to what they return and which are cheap. You can set a budget per workflow that reflects its value rather than the headcount of the team that happens to own it. And you can govern an agent fleet the same way, because every agent runs a workflow, and a workflow has a budget, an owner, and a stop condition. This is where the agent badge earns its keep on the finance side. An agent with a named identity scoped to one workflow is also an agent whose spend lands on that workflow's line.
How do you actually do this?
Tag every token to a workflow at the point of spend, not a person at the point of login, and make the tag mandatory before the call goes out.
The build is smaller than it sounds and it's mostly plumbing. Every request your systems send to a model carries a workflow identifier alongside the credential, and the routing layer you own refuses requests that don't have one. The identifier flows through to the usage logs and the invoice, so spend aggregates by workflow automatically. Individual experimentation gets its own identifier, which is how you keep the discovery budget visible without pretending it's production. Service accounts stop being a black hole, because a service account can run twenty workflows and each one now shows up separately. None of this requires a platform purchase. It requires a routing boundary you already should have, which we've argued for since the model socket, and a rule that nothing crosses it unlabeled.
What should you do about the seat reflex?
The question you're really asking is why your AI bill keeps surprising you when every person is under a cap. The answer is that the bill was always the workflows, and you haven't been counting them.
Pull last month's AI spend and try to assign every dollar to a workflow. Most companies can assign the individual seats and almost nothing else, and the unassignable remainder is the seat reflex in dollar terms. Then change the unit. Put a workflow tag on every call that crosses your boundary, starting with the agents and pipelines that run under shared credentials, because that's where the untracked spend concentrates. Give each workflow a budget sized to its outcome, an owner who sees the number weekly, and the same stop condition you'd give any standing agent. Leave the individual caps where they are if you must, but understand what they do. They govern the ten percent of spend you could already see. The tag governs the rest.
If you want your AI spend counted by the work it does instead of the people it's assigned to, start with an AI Blueprint or reach us at contact@theyor.com