What an AI agent actually costs to run

The short answer: four lines, roughly in ascending order of how much people underestimate them — the platform, model usage, setup, and ongoing human attention. Model usage is the volatile one and it is driven by the shape of the job, not its importance: an agent that reads twenty web pages to answer one question costs an order of magnitude more than one that classifies an email, and both look identical on a feature list.

Line 1: the platform

Predictable, and the one people fixate on. Worth checking only for how it scales: per seat, per usage, or per deployment. A seat is a person, and an agent is not one, so per-seat pricing gets strange fast once agents outnumber staff. We wrote that comparison out in per seat, per usage or per deployment.

Line 2: model usage, which is where the variance lives

Three things drive it, and only the third is obvious.

How much the agent reads. Every document, web page, search result and prior message in the conversation is input. An agent that researches a company by reading its site is processing tens of thousands of words per task. An agent that triages an email is processing a few hundred. Same platform, two orders of magnitude apart.

How many steps it takes. An agent that calls a tool, reads the result, decides, calls another and writes a reply has made several model calls for one task, and each one carries the accumulated context. This is why "it just answers questions" and "it handles the whole workflow" are different cost categories rather than different feature tiers.

Which model. The obvious lever, and the one to use last. Routing the easy majority of tasks to a fast model and reserving the capable one for the hard minority is usually worth more than any prompt optimisation — see multi-model agents.

The practical consequence: estimate per job, never per agent. "Five agents" tells you nothing. "One agent doing 200 email triages a day and one doing 20 deep research tasks a day" tells you almost everything, and the second one is the expensive one despite being a tenth of the volume.

Line 3: setup, paid once per job

Deciding what the job is, connecting the systems, writing down the procedure, and the two weeks of reviewing every output while you find out what context the agent was missing. Mostly time rather than money, and mostly front-loaded: the second agent on the same systems is dramatically cheaper than the first, because the connections and the conventions already exist.

The part that gets skipped is the reviewing fortnight, and skipping it does not remove the cost — it moves it into month three as an agent that has been confidently wrong in a way nobody checked.

Line 4: attention, which is the one that decides whether this was worth it

Someone has to look at what the agents did. Not constantly — but a weekly fifteen minutes per agent, and a named person who notices when answers drift.

This is the line that is never in a business case and is frequently the largest. It is also the one that determines outcomes: the deployments still running a year later all have this person, and the ones that quietly stopped being trusted all lacked one. Treat it as a real cost, budget it explicitly, and the rest of the model holds.

The costs that appear in month two

• A retry loop. An agent that fails a step, retries, and fails again is billing you for every attempt. Worth a cap and an alert rather than a discovery.
• Context growth. A long-running conversation carries its history into every call. Good platforms bound this; it is worth confirming yours does.
• An enthusiastic scheduled task. Hourly is eight times daily. A monitoring job set to run more often than the thing it monitors changes is the most common source of a surprising invoice.
• A connector that broke quietly. Not a cost in itself — but an agent retrying against a dead connection is, and nobody notices until the bill arrives. Why MCP connections keep breaking.

Building an estimate you can defend

1. List the jobs, not the agents. For each: how many times a day, how much it reads, how many steps.
2. Run the noisiest job for a week and measure it rather than modelling it. One week of real usage beats any spreadsheet.
3. Multiply by the others, then add half again. The half is for the retries, the context growth and the job you have not thought of.
4. Add the human line explicitly — fifteen minutes per agent per week, plus the setup fortnight.
5. Set a hard cap anyway. Not because you expect to hit it, but because the failure mode of an unbounded agent is financial rather than technical.

Comparing this against a person

Only once you have the four lines, and only for the fraction of a role that is actually automatable. Doing it with a subscription price on one side and a salary on the other produces a number that is wrong in both directions — that comparison, with the official US employer-cost figures, is in hire or deploy.

And if you are weighing this against running the stack yourself, the infrastructure line is the small part of that too: what self-hosting really costs.