Most AI agent proposals contain a price: a licence per seat, a price per million tokens, or a monthly platform fee. A finance director reads that number, multiplies it by twelve and approves it. Six months later the real spend is higher, and nobody can explain exactly why. The licence was never the whole cost.
This article is for the manager who has to put a yearly figure for an AI agent into a budget. It assumes you already know which process the agent will handle. If you are still choosing a supplier, start with how to choose an AI agent vendor. Here the question is how to count the full cost and plan money you will not have to explain away.
1. Count five groups of cost
A budget built from five groups catches most of what a price list leaves out.
- Build and integration (mostly one-off). Connecting the agent to your systems, preparing the knowledge it uses, setting up access rights, testing, and the time of your own people during the project.
- Usage (recurring). Model calls, platform fees, hosting, search or storage components, and any per-seat or per-case charges.
- People time (recurring). Review of the agent's output, correcting errors, handling escalations, and an owner who spends part of the week on the agent.
- Maintenance and change (recurring). Updating instructions, sources and rules when the business changes, re-testing after a model or vendor update, and responding to incidents.
- Exit (one-off, often forgotten). Export of data and configuration, switching to another supplier, and a manual fallback if you need to retire the agent. The related questions are covered in vendor contracts, SLAs, liability and exit.
Write each group as a line in the same sheet, with an owner. A line without an owner will be the one that overruns.
2. Plan usage as a range, not a single number
Usage cost depends on two things you only partly control: how many cases arrive, and how much text or data the agent processes for each. Public price lists for model APIs charge per unit of input and output, and the prices differ by model, so the same agent can cost very different amounts depending on the model and how it is configured (see for example the OpenAI API pricing page and the Anthropic pricing page). Prices change, so check the current page when you build the sheet and note the date.
Use three cases:
- Low: the volume you saw in the pilot, with short, simple cases.
- Expected: the volume you plan for after rollout, with a realistic mix of simple and long cases.
- High: a busy season or a wider rollout, plus the case where the agent retries or reads long documents more often than you assumed.
Ask the team to measure the average cost per handled case during the pilot, then multiply by each volume. If you cannot measure it yet, say that the figure is an estimate and plan a check after the first month. Limits on spending and loops, described in cost controls, budgets and loop limits, keep the high case from becoming an unplanned one.
3. Give people time a price
Review and rework are the cost that most proposals leave out. If a person checks every output for two minutes and the agent handles 3,000 cases a month, that is 100 hours a month, whatever the licence costs. Multiply the hours by an internal hourly cost and put the result in the budget, next to usage.
Include three kinds of time:
- Review of correct output. Approval steps are a deliberate control, but they cost time. Decide which cases need them and which can be sampled instead.
- Rework of wrong output. Estimate the share of outputs that need correction, and how long a correction takes compared with doing the case from scratch.
- Ownership. Somebody has to read reports, talk to the supplier and decide what changes. Even half a day a week is a real cost.
The share of review should fall as the agent proves itself, but do not plan on it. Budget for the current level and treat any fall as a saving you report later, as shown in a one-page scorecard for leadership.
4. Plan for maintenance, not only for launch
An agent is not a finished product. Your products, prices, policies and forms change, and the agent's instructions and sources have to change with them. Suppliers update models, which can change behaviour even when you touch nothing. Allow time for regular review of samples, updates to sources, regression tests and incident handling. The routine is described in maintaining an AI agent after launch.
A simple way to budget it: estimate the monthly hours for the owner and a technical person, price them, and add a line for the supplier's support tier if you need one.
5. Add a reserve and say what it is
A reserve covers what you cannot foresee: extra integration work, a longer pilot, a price change. Many organisations add a contingency to project budgets. A range of 15 to 25 percent is a working rule of thumb we use for first agent projects, and it is our own assumption rather than a published benchmark. Choose your own figure, write down why, and release the reserve only with the owner's explanation.
6. Treat the budget as managed spend
The discipline of cloud financial management applies well here. The FinOps Foundation describes FinOps as a practice that brings finance, engineering and business together so that teams take ownership of their usage and make trade-offs with a shared view of cost (FinOps Foundation). Cloud providers publish similar guidance on cost optimisation, such as the Google Cloud Architecture Framework and the Azure Well-Architected cost pillar. You do not need a FinOps team for one agent. Three habits are enough:
- Tag or separate the agent's costs so that they can be seen on their own.
- Review the actual spend against the plan every month for the first quarter, then every quarter.
- Assign a named person who can explain a difference between plan and actual.
7. Compare the cost with the right alternative
A total cost figure means little on its own. Compare it with what happens without the agent: the current cost of the process from your baseline, the cost of hiring or outsourcing the same capacity, and the cost of doing nothing, such as slower response times or a growing backlog. If you cannot show a credible comparison, the honest answer may be a smaller pilot rather than a bigger budget. Methods for the baseline are in how to measure whether AI agents really save work.
Do not promise a payback period you cannot support. Show the range from your low and high cases, and say which assumptions it depends on.
8. Ask these questions before approving the figure
- Does the budget include all five groups: build, usage, people, maintenance and exit?
- Is usage shown as low, expected and high, and is each based on pilot data or labelled as an estimate?
- Is review and rework time priced, and is the volume it is based on stated?
- Is there a named owner for each line?
- Is the reserve explained, and who can release it?
- Is there a review date, with a condition under which the spend is reduced or stopped?
- Is the comparison with the current process based on a measured baseline?
A short checklist
- Write the five cost groups into one sheet with owners.
- Measure the cost per case in the pilot and plan three volume cases.
- Price review, rework and ownership time.
- Add maintenance hours and an exit line.
- Set a reserve and explain it.
- Approve the budget for a fixed period, with a review date and a stop condition.
A budget that shows the full picture is easier to approve and much easier to defend. When the first quarterly review comes, you will be explaining a plan you made, not discovering a bill.
