Resources · automationpricingagencies
What it costs to keep an AI agent running each month
Keeping one production AI agent running usually costs a small, boring amount most months. For a single-workflow agent, that is tens of dollars in AI calls plus an hour or two of a human's time, so under a hundred dollars in a quiet month. The number you cannot plan around is the spike: the month a connected tool changes something and someone has to notice, diagnose, and re-wire it.
That is the honest shape of it, and almost nobody publishes it. Search this question and you get dev-agency posts guessing "$500 to $10,000 a month" with no line items and no methodology. Those are people who have never run one for a year. Here is what a real month looks like.
What are you actually paying for each month?#
Five line items, and only two of them cost real money in a normal month. This is the table to hand your vendor and make them fill in for your build.
| Line item | What it covers | Typical band | What makes it spike |
|---|---|---|---|
| AI and API calls | The model requests the agent makes to read your data and decide what to flag | Tens of dollars a month for one workflow | More runs per day, longer context, more messages to read each run |
| Human upkeep | Someone edits a rule, reads the flags, moves a threshold from 21 days to 14 | An hour or two a month | A busy season, a rules doc that needs a rewrite, a new signal you want caught |
| The watching layer | A small check that reads the agent's own logs and flags when it goes quiet or its volume halves | Near zero on its own, folded into upkeep | Almost never, and that is the point of building it |
| Tool subscriptions | The paid seats the agent needs: the automation platform, any API plan | Flat, set by the vendors, not by us | A vendor raises prices or pushes you into a higher volume tier |
| Break-fix | The month a connected tool changes an export format, an API limit, or a login flow and the agent has to be re-wired | Zero most months, a few hundred dollars when it happens | A tool ships a redesign, deprecates an API, or a model version gets retired |
The first four are predictable. You can budget them to the dollar. Break-fix is the one that makes the yearly number hard to guess, because you cannot know in advance which month one of your tools will change underneath you. You only know that over a year, one or two of them will.
Why is the monthly bill spiky instead of flat?#
Because the agent's cost comes from the tools it is wired into, and you do not control when those tools change. A month where nothing changes is cheap. The expensive months are the ones where a tool changes: Seller Central redesigns a report, a CRM changes its export columns, or a model version gets retired. Then someone has to spot the drift, work out what moved, and reconnect the agent to the new shape of the data.
This is the whole reason flat monthly retainers exist, and why they are not a markup. A retainer charges you the average of a spiky year so that the month a tool breaks does not land as a surprise invoice. You are buying a flat line over a jagged one. A shop that quotes you a single flat number with no mention of what happens when a tool changes has either priced the spikes in quietly, or has never lived through one. Ask which.
What does maintenance actually look like on an agent we run?#
Most months, nothing happens. The agent runs each morning, reads what it reads, posts its flags to one Slack channel, and the only human cost is the person glancing at those flags. We do not touch it. The bill is the AI calls and a few minutes of attention.
The work comes on the spike months, and it always starts with the watching layer catching drift before the client does. Every agent we run reports what it did to one channel, and a second, dumber check reads those reports and raises a hand when something looks off: this one went quiet three days ago, this one's volume halved, this one started erroring. That flag goes to us, not just to the client, because a broken automation that fails silently is the expensive kind. The failure itself is cheap to fix. What costs you is the weeks it ran wrong before anyone noticed.
Then a person diagnoses it. Usually a connected tool changed a screen or an export, and the fix is to re-map the agent to the new shape and re-run it against recent history to confirm it behaves. We move the approval gate back to manual for a run or two while we watch, then hand it back to automatic. That is the few-hundred-dollar month. It is real work, it is rare, and it is why the run number in what an AI build costs is a monthly line and not a footnote.
What should you ask your vendor about month three?#
The build works on day one. The question is what happens in month three when a tool changes underneath it. Ask these before you sign, not after.
- What happens the first time one of my tools changes its export format or its login screen? Who notices it, and how fast?
- Is upkeep time inside a flat retainer, or billed by the hour when something breaks? Get the answer in writing.
- What did last month's bill actually look like on a system like mine? Ask for the line items, not a range someone made up.
- Who gets the alert when the agent goes quiet, and what is the response time? If the answer is "the agent will tell you," ask what tells you when the agent itself stops.
One honest caveat before you budget any of this. If a workflow only runs a few times a month, the break-fix exposure can cost more than doing it by hand. Ask the vendor to tell you when that is true for you, and to say so even when it costs them the build.
The short version: keeping an agent running is cheap and boring most of the year. Budget for the two months it is not, insist on a watching layer that catches drift for you, and treat any vendor who cannot name the break-fix line as someone who has not run one long enough to have hit it.