Somewhere around April, a pattern started showing up in finance reviews that nobody had budgeted for. Companies that had approved reasonable AI spend for the year were already through it. Uber, by one account, maxed its AI budget four months into 2026, driven largely by engineers running coding agents. It wasn't alone. By the halfway mark of the year, plenty of enterprises were tracking three times over their token budgets, with the second half still to pay for.
The reflex in most boardrooms is to treat this as a pricing problem. Negotiate the contract. Cap the licenses. Wait for the per-token cost to fall, which it will. That reflex misreads what is happening. The bill isn't large because the model is expensive. It's large because of what the model is being asked to do, over and over, on top of processes that were never built to be run at machine speed.
A meter on top of the dysfunction
Agentic systems earn their reputation by not giving up. They reason, they loop, they retry. That persistence is the point when the underlying task is clean. It's also the mechanism by which a weak workflow turns into an open-ended expense.
J.R. Storment, who runs the FinOps Foundation, put the technical version plainly: adding reasoning and loops and retries "created a lot of loops and an extravagant amount of token usage." Read that as an operating statement, not a technical one. Every retry is the system trying to force its way through a step that a person, working the same process, would have flagged, escalated, or refused to do. The agent doesn't have that judgment. So it loops. And each loop is metered.
This is why the cost surprises land at quarter-end rather than at deployment. The dysfunction was always in the process: undocumented handoffs, unclear ownership, decisions no one had defined. It used to be absorbed silently by people working around it. Point an agent at the same process and the workaround becomes a line item. Bad workflow plus AI doesn't only produce faster bad outputs. It produces a meter that runs while it does.
The cost structure makes this worse than it looks. At Shutterstock, the CTO found that about 75% of token consumption was output, not input. The expensive part is the volume the system generates, the thing that multiplies when logic is loose. A routine FinOps review at the same company surfaced a quarter-million dollars in vendor commitment that would have gone unused, and a jump from 50 to 750 ChatGPT licenses inside a single month. None of that is a model-pricing story. It's a governance-and-ownership story wearing an AI label. And it scales: Goldman Sachs projects global token usage will multiply about twenty-four times by 2030.
What moving fast bought at Bristol Myers
The counterexample everyone points to is Bristol Myers Squibb, and it's worth being precise about why it worked, because the wrong lesson is easy to draw.
BMS took its RFP cycle from six-to-nine months down to under thirty days, ran more than a billion dollars of sourcing through an AI procurement platform in the first year, and expanded RFP volume roughly tenfold with about half the resources. Its executive on the project, Rhonda Griscti, is blunt about method: "If you wait for perfect data, you'll never get started." Deploy on a data lake, clean as you go, don't let the pursuit of a spotless input hold the whole thing hostage.
That sounds like a mandate to move fast on imperfect data. It isn't. Or rather, it only reads that way if you ignore the shape of the workflow. Procurement decisions are bounded, reviewable, and reversible. A bad RFP output gets caught before a contract is signed. Griscti could run on imperfect data because the process had containment: a place where a wrong answer surfaces before it costs anything. The speed wasn't the achievement. The containment was. Copy the speed into a process that lacks it, and you don't get a thirty-day cycle. You get flawed evaluations and mispriced work produced faster than anyone can review them.
The number nobody owns
The data-trust picture explains why so few operators have that containment. An RGP report late last year found only 10% of CFOs fully trust their data quality, and more than a third name data trust as their single biggest barrier to AI return. Info-Tech's research points the same direction from the governance side: organizations with a board-governed AI strategy are about three times more likely to pull real value out of the work. Info-Tech's Tom Zehren also offers the discipline that gets skipped in the rush to cut cost: "Don't fire half your team, don't reduce capacity massively in IT. Repurpose it." The instinct to buy an agent as a headcount substitute is how an unowned process becomes a metered one.
Because that's the trap. An operator who buys agents to paper over an undocumented, unowned workflow hasn't cut cost. They've converted a hidden process problem into a visible, per-token one. Now they pay the model to retry its way through the same dysfunction, and find out only when finance flags the overage.
So the question in front of any executive approving this spend isn't the one the vendor invites. It isn't "what does the model cost per token?" That number is going to fall regardless. The real question is whether the workflow this agent runs is clean enough that every loop it takes is one you would have wanted a person to take. And when the monthly bill triples, who owns that number?
If you can't name the owner, you don't have a pricing problem. You have execution debt, and it's already accruing interest.
Frequently asked questions
Why does an AI agent run up such a large token bill?
Because agentic systems reason, loop, and retry until they reach an answer. When the underlying workflow is undocumented or unowned, the agent has to force its way through steps a person would have flagged or skipped, and every one of those retries is metered. The bill reflects the state of the process, not the price of the model.
Isn't a runaway token bill just a pricing problem I can negotiate down?
Per-token prices are falling and will keep falling, so negotiating helps at the margin. But the bill is large because of how many times the model is asked to work through a broken process, not because each call is expensive. Fix the workflow and the same work costs a fraction. Leave it and the meter keeps running, whatever the rate.
Bristol Myers deployed on imperfect data and still won. Doesn't that prove "move fast" works?
It proves containment works. BMS could run on imperfect data because procurement decisions are bounded, reviewable, and reversible. A wrong output surfaces before a contract is signed. The speed wasn't the achievement. The control around it was. Copy the speed into a process without that containment and you get mistakes at machine speed.
How do I know whether a workflow is ready to hand to an agent?
Ask whether every loop the agent would take is one you would have wanted a person to take. If the steps, handoffs, and decision owners are defined, the agent's persistence works in your favor. If they aren't, the agent industrializes the dysfunction, and you learn about it at quarter-end when finance flags the overage.
Who should own AI token costs inside the business?
A named owner tied to the workflow, not "IT" in the abstract. The organizations that get value from AI treat cost as a governed number with an owner, the way mature teams govern cloud spend through FinOps. If no one can say who owns the bill when it triples, the process isn't ready to scale, whatever the model costs.