What is the biggest cost of running AI agents?

Kaykas-Wolff argues the largest cost is usually the human time spent reviewing and fixing agent output, which never appears on the model invoice. A draft that needs a full rewrite costs the same in model fees as one that is sent as is, so a token dashboard can make an agent look cheap while it consumes hours of salaried review time.

How should executives measure the cost of an AI agent?

By cost per accepted output. Add the agent's monthly model spend to the reviewer's minutes at a loaded hourly rate, then divide by the outputs a human used without rewriting. A two-week tally from the reviewer of outputs used versus rewritten, plus a rough estimate of daily checking time, is enough to start.

Who should own the budget for AI agents?

The person who owns the agent, including the review time it creates. Kaykas-Wolff warns that a central AI budget line teaches departments that agents are free. The CFO keeps the monthly reconciliation and the spend caps, and an owner can raise a cap with a written reason. The CEO's own agent spend goes to the CFO for review.

Can the AI bill help find shadow agents?

Yes. Every agent that calls a model leaves a charge somewhere. Reconciling those charges against the fleet register each month surfaces any charge without a named agent and owner, which is either a shadow agent someone built without approval or a retired agent that never actually stopped running.

What should a board see about AI agent costs?

One line per quarter: total agent spend including an honest estimate of review hours, set against what the fleet produced. If management can only report the model invoice, Kaykas-Wolff says directors are looking at the smaller number and calling it the cost.

AI Agent Costs: Price the Review Time, Then Decide Who Owns the Bill

Nearly every guide to AI agent costs opens on the same scene. An agent gets stuck in a loop overnight, calls a model a few thousand times, and finance learns about it when the invoice arrives. That happens. I once had a retired agent of mine keep waking up on its old schedule for months.

But the runaway loop is the dramatic cost, and dramatic costs get caught. The one I worry about is quieter and shows up every day, in a column nobody puts on the dashboard.

What the cost guides get right

The pages that rank for this are mostly written by FinOps and API gateway vendors, plus a long InfoWorld piece on deploying agents, and they agree on a short list. Give each agent its own budget, because one pooled company limit cannot tell you which workflow blew through it. Send routine work to cheaper models and save the expensive one for hard calls. Detect loops and kill them. Several frame the problem as two executives staring at the same agent from different chairs, a security chief worried about exposure and a CFO worried about spend.

I run most of those controls in the fleet I use at Visiting Media, and I would not argue with any of them. The per-agent cap least of all. In my post on writing an AI risk appetite statement I said weekly spend per agent should be capped inside the system, so the ceiling stops the money before it moves. I still think that.

All of these guides are measuring the model bill, though.

The line that never reaches the invoice

Take an agent that drafts customer follow-ups. Each draft costs a sliver of a cent in model fees. Then a person reads it. If they fix two sentences and hit send, the review took a minute. If the agent mixed up the account and the person rewrites the whole thing from the top, the review took longer than writing it fresh would have, and the model charge for that bad draft is identical to the charge for the good one.

That review time is salaried. Nobody logs the twelve minutes spent untangling an agent’s summary of a board deck, so the minutes dissolve into a busy week, and the agent keeps showing up on the cost dashboard as a bargain. My bet is that in most fleets, mine included, the people checking the work cost more than the models doing it.

A cheap agent that needs a babysitter is an expensive agent.

Price agents by the output somebody used

The number I care about is cost per accepted output. Add up an agent’s model spend for a month. Add the reviewer’s minutes at a loaded hourly rate. Divide by the outputs a human actually used without rewriting. Discarded drafts land in the numerator and never in the denominator, which is exactly how they should hurt.

None of this needs software. Ask whoever reviews an agent’s work to keep a tally for two weeks, outputs used versus outputs rewritten, with a rough daily guess at minutes spent checking. It will be imprecise. It will also be the first honest number you have.

The ranking that comes out of it rarely matches the token dashboard. A research agent that runs on the priciest model can turn out to be the cheapest per useful answer, because nobody edits what it hands back. A drafting agent on a small model can look like a steal while eating an hour of someone’s afternoon. This is also where routing advice goes wrong when it is followed blindly: move a task to a cheaper model, watch the invoice fall, and miss that the rewrite rate went up. I made a similar argument in change control for model upgrades, where a cheaper model that needs retries is the more expensive one.

The bill is also an inventory

The cost guides treat the invoice as something to shrink. I use it as something to reconcile. Every agent that calls a model leaves a charge somewhere, usually on an API key and occasionally on somebody’s expense report, and once a month those charges should be matched against the fleet register line by line. A charge with no named agent and no named owner behind it is one of two things. Either it is a shadow agent someone built without asking, or it is a retired agent that never actually stopped, which is the exact failure I wrote about in retiring an AI agent. Mine was found because a machine ran out of process headroom. A reconciled bill is a less embarrassing alarm.

Who owns the bill

Whoever owns the agent owns its cost, review minutes included.

A central AI budget line feels tidy, and I understand why finance teams like it. It also teaches every department that agents are free, because the money never leaves their own budget. When an agent’s cost lands next to the work it was supposed to replace, its owner starts asking whether it is worth keeping. That is the question you wanted asked anyway. The CFO keeps the reconciliation and sets the caps, and an owner can raise a cap with a written reason, on the same tiers as the delegation of authority matrix.

The CEO’s own agents follow the same rule, with one change. Their bill goes to the CFO for review, alongside the grants I described in AI agent access reviews, because nobody should approve their own spending.

A note on caps

A spend cap is the cheapest kill switch you will ever install, since an agent that cannot pay for its next model call stops, whatever else is wrong with it, and that is worth remembering the night you are reaching for the incident playbook.

What the board should see

One line. Total agent spend for the quarter, with an honest estimate of review hours added in, set against what the fleet produced. That is the consolidated spend row I recommended for the AI page in the board pack, with the hidden half finally included. If management can only report the model invoice, directors are looking at the smaller number and calling it the cost.

Start with your favorite agent

Pick the agent you are proudest of, the one you describe at dinner as practically free. Ask its reviewer for the two-week tally and do the division. If it still looks like a deal, good, you now have a number you can defend in front of your CFO. If it does not, you have found the first thing to fix, and I would bet the fix has nothing to do with which model it runs on.

This article is part of the Executive Leadership cluster, focused on board governance and the operating discipline required to run AI systems responsibly at the executive level.