A support agent on one of our client projects costs between 10 and 15 US cents per question. We measured that in September 2026, on Anthropic's Haiku 4.5, with the agent reading the client's documentation and nothing else yet. At ten thousand questions a month the bill lands between 1,000 and 1,500 dollars, which buys less than a week of a support person's time in most of Europe.
Hold on to that number, because it is the easy half of the calculation. It is also the half that ends up in the board deck.
What does an agentic AI pipeline actually cost?
An agentic pipeline costs its tokens plus everything the tokens leave out: retries, the time a human spends reviewing the output, supervision when a run goes sideways, rework when the output was confidently wrong, and the long tail of owning software nobody on the team wrote. The token bill is the visible part. Across every engineering decision we have tried to price, the visible part is usually 10 to 30% of the total.
That ratio is the spine of Bridging Code and Capital, the book on engineering economics we are publishing this autumn. We worked it out for bugs, bad hires, rewrites and manual work, where the full cost came to six and a half times the direct labour everybody calculates. Agents are the first case where the visible part arrives as a neat invoice, which makes it very tempting to stop counting there.
Here is one cycle of a coding agent on a client project in September. The pipeline picked up the review comments on a pull request, made about 45 tool calls in ten minutes (five of them spent waiting for the test suite), addressed six comments and flagged one it could not handle. Cheap, fast and genuinely useful. It also ended with the words "needs a human", and that human is not on the invoice.
How does that compare with an engineer doing the same work?
Compared honestly, the agent wins on unit cost and loses on predictability. An engineer is a fixed cost you forecast once a year in a hiring plan. An agent pipeline is a variable cost that scales with usage, retries and the size of the codebase it has to read, and it can double in a month without anyone signing off on it.
| Engineer | Agent pipeline | |
|---|---|---|
| How you pay | Salary, fixed | Tokens, per use |
| What drives the cost | Headcount | Usage, retries, context size |
| Who forecasts it | Finance, in the hiring plan | Usually nobody |
| How it fails | Slowly and visibly | Quickly and confidently |
| Where the hidden cost sits | Onboarding, management, attrition | Review, supervision, rework, ownership |
The comparison is getting less theoretical by the quarter. Gartner expects AI coding token costs to match or exceed a typical engineer's monthly salary within two years, and reports single developers already spending 20,000 dollars in a month (CIO, June 2026). Their benchmark is a global average salary of 2,000 dollars a month, so the crossover arrives later for a team in Ghent or Berlin. It still arrives.
The hidden costs on the agent side are measured too. Faros AI followed 22,000 developers through two years of telemetry: throughput went up, and so did incidents per pull request and median review time, by a wide margin (Faros AI, 2026). As we've put it earlier, generation got faster and review did not.
Why is token spend harder to forecast than cloud spend?
Token spend is harder to forecast than cloud spend because the price has not settled, and it moves in both directions at once. Cloud got cheaper and more predictable year after year. AI is getting cheaper faster than cloud ever did, whilst the amount a team uses grows faster still.
If you lived through a cloud migration in the 2010s, this will feel familiar. Moving servers to AWS turned a capital expense into an operating one, and a few quarters later someone in finance opened a bill nobody could explain. A whole discipline, FinOps, grew out of that surprise. Token spend is the same story on fast forward.
In May we predicted the end of the all-you-can-eat buffet. So far, the buffet has only grown. Anthropic doubled the five-hour Claude Code limits on its subscriptions that same month. When it announced that automated usage would move off those plans onto separate credits, it paused the change on the day it was due in June. Claude Opus 5.5, released on 22 September, lists at 20% below Opus 5, and Anthropic says it costs 40% less to run because it also needs fewer tokens for the same task. Half of that saving never appeared on a price list, and that is the forecasting problem in miniature: the cost of a task moves between releases for reasons a rate card does not show. None of it makes the bill easier to plan either. A cheaper model is a reason to run more agents, and the monthly total follows usage. And the flat rate is still a subsidy the vendor has shown it can announce the end of with a month's notice.
There is a quieter trap as well. Engineers without an API budget route their agents through personal subscriptions, which keeps the bill flat and the usage invisible. You cannot forecast what you cannot see.
What should investors ask about AI spend in due diligence?
Investors should ask what the AI tooling costs per unit of work delivered, who owns that budget, and what happens to margins if the price doubles. A team that can answer all three has treated AI as an engineering decision. A team that answers with a velocity chart probably has not been asked yet.
In the technical due diligence we run for investors, the AI question has moved on from whether a team uses it at all. The useful follow-ups are concrete. What does the monthly AI bill look like, and is it growing with revenue or ahead of it? How much of it sits inside the product, where it hits gross margin, rather than in engineering tooling? Who reviews what the agents produce, and how much of their week does that take? If the flat-rate plans disappeared tomorrow, what would the same usage cost? Those answers now belong in a due diligence report next to the hosting bill.
We are still working out some of our own answers, from who pays for the tokens to how you bill for an engineer running three agents in parallel. That is exactly why we ask.
Is the agent still worth it?
Often, yes. For another client, people were researching up to 500 companies a week by hand, visiting four sites for each one, and a prototype now does each company in under three minutes. No salary comparison survives that.
The point is to compare the full cost with the full cost. Tokens against salary is the version that fits on a slide. Tokens plus review, supervision, rework and ownership, set against salary plus management, onboarding and attrition, is the version that tells you whether you are buying capacity or borrowing it. We argued earlier this year that AI did not change the economics of software engineering, because producing code was never the expensive part. What it changed is where the bill turns up, and who gets to read it first.
