There are three numbers. One is the cost to build, one is the cost of each run, and one is what the agent may commit for you. A quote usually contains the first, may mention the second, and rarely names the third. When an agent goes wrong, the third number does the most harm.
Ask a vendor what an agent costs and you will get a build fee and a monthly figure. Both are real. Neither tells you the cost in month seven. The two costs that grow rarely appear on a slide.
An agent has three costs
- /01
Build. Paid once. It covers scoping the process, connecting its systems, and the first weeks of fixes while the agent learns your exceptions. It is the number on the quote, and it stops.
- /02
Run. Paid each time work arrives. Model tokens, retrieval, retries, and the human minutes spent on whatever the agent held. It grows with your volume, as it should. It also grows with messy inputs, which you can fix.
- /03
Commitment. Paid each time the agent decides something in your name: a refund approved, an invoice matched and posted, an order released. This is money the agent moves on your behalf. It is rarely on a quote, and it is usually the largest number.
Per seat pricing is a warning sign
Seats price software a person sits in front of. An agent does not sit in front of anything. Per seat, you pay the same for forty items or four thousand. Either the vendor charges too much now or loses money as you grow. Ask for a per run price and a volume band. If the answer is that per run is hard to calculate, that tells you their system does not measure its own costs.
What moves the per run number
- /01
Document length. A forty page contract costs several times what a one page invoice costs, and the gap stops being linear once retrieval is involved.
- /02
Retries. A system that retries three times before succeeding costs three times what its headline price implies. Ask whether retries are billed and whether you can see them.
- /03
Prompt weight. Prompts grow. Each time somebody fixes an edge case by adding a paragraph of instruction, future runs pay for that paragraph.
- /04
Held runs. A hold is cheap in tokens and expensive in minutes. If one run in ten waits six minutes for a person, that is a real cost. It is also the cost most worth cutting.
| Line | Who watches it | What makes it grow |
|---|---|---|
| Model tokens | The vendor, on their margin | Longer inputs, larger models, growing prompts |
| Retrieval | Usually nobody | More documents in scope, weaker search |
| Retries | Nobody, until the invoice | Flaky integrations, ambiguous inputs |
| Human minutes on holds | You, in headcount | A threshold set too cautiously |
| Correction after the fact | You, in rework | No threshold at all |
The commitment is usually the largest cost
Running costs are usually small. What the agent commits can be large. An agent triaging claims might cost cents to run while being trusted to approve settlements worth thousands. Treat them as separate budgets, or a small project can turn into a large incident.
So they get separate limits. A compute limit stops a run that is burning money for no reason. A commitment limit stops a run that is about to promise something you would not have promised. At either limit the right behaviour is to hold, keep the work already done, and ask a named approver. If a limit throws work away, people find ways around it.
GatehouseSpend limits sit in the signed manifest, and Gatehouse, Surehand's control plane, enforces them while the agent runs. See Gatehouse
Add up the first year before you sign
Before you sign, add up the build and the run cost at your real volume. Add the staff time at your expected hold rate and the work to keep integrations running. Then estimate one bad month if nobody watches for drift.
| Ask for | Why it matters | A weak answer sounds like |
|---|---|---|
| A per run price | It is the only number that survives growth | It depends on usage |
| What is billed on a retry | Retries are the hidden multiplier | Retries are rare |
| Your expected hold rate | It converts directly into headcount | The system is very accurate |
| A commitment limit, enforced | It is the number that stops a bad month | We monitor approvals closely |
| The price at three times volume | You will get there, or you did not need this | We can revisit at renewal |
How Surehand prices a deployment
The proposal states the price. The signed manifest sets a compute limit and a separate commitment limit before anything is deployed. The teardown is scoped and priced in writing before it starts. It measures one process, and you keep the document.
Keep reading
How to set a spend limit on an AI agent
Set it in money rather than tokens, at three levels, enforced in the system. How to derive the number from the process you are replacing.
What should you ask an AI vendor before signing?
Twelve questions that separate a system you can run from a demo you cannot govern. Ask us the same ones.
