Budget per capability, per case, and count the people as well as the computing. The running cost of an AI agent is model usage and platform fees, plus the time people spend approving held items, re-certification, monitoring, records and change testing, plus a reserve for incidents. Forecast volume, price each case, cap spend per agent and review monthly.
What does an agent budget cover, and who owns it?
Agent budgets are often built from the model price alone, which is usually not the largest line. An agent working at depth 2 drafts and a person sends, so every case still carries some staff time. Every capability needs testing at certification and after changes. Records have to be stored for years. The budget that matters is the cost of a case done well, compared with the cost of the same case without the agent.
The process owner owns the budget for their agents, because the savings and the risks land in their process. Finance sets the method and tracks spend. The agent operations lead provides unit costs for shared items such as the platform and the operations team. The full path from pilot to wider use, where costs change shape, is in deploying an AI agent: train, shadow, certify, widen.
Which cost lines should the budget include?
- Model usage: charges per call or per volume of text processed, by capability.
- Platform and connectors: licences, hosting, and maintaining integrations as connected systems change.
- People on held items: approver time per held case, multiplied by expected holds.
- Agent operations: the team that monitors, triages refusals and runs changes.
- Certification: replaying cases and reviewing results at each renewal and after significant changes.
- Assessments: data protection impact assessments and, where they apply, fundamental rights impact assessments; FRIA vs DPIA for AI agents explains how to run them together and avoid paying twice.
- Records: storing the decision record for as long as you must keep it.
- Incident reserve: a sum for suspension, manual fallback and remediation.
What are the steps?
- Person, process owner: forecasts volume per capability: cases per month, with peaks.
- Agent, depth 1: measures unit cost in shadow: model usage per case, by capability, from real runs on past cases.
- Agent, depth 1: estimates held volume from the shadow run: how many cases per capability would be held at the planned depth.
- Person, operations lead: prices people's time on held items and agent operations.
- Agent, depth 2: drafts the budget by capability and cost line, with a comparison against the current cost of the same work.
- Person, process owner: approves the budget, including the incident reserve.
- Person, finance: sets spend caps per agent with the model provider and platform, with alerts at agreed levels.
- Agent, depth 1: reports actual cost per case each month, by capability.
- Person, process owner: reviews and adjusts monthly in the first quarter, then quarterly.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Forecast volume | Nothing | none | Setting volume assumptions |
| Unit cost | Measures usage per case in shadow | 1 | Nothing |
| Held volume | Estimates holds per capability | 1 | Nothing |
| Price people | Nothing | none | Costing staff time |
| Draft budget | Builds budget by line and capability | 2 | Nothing |
| Approve | Nothing | none | Owner's decision |
| Caps | Nothing | none | Setting hard limits |
| Report | Reports actual cost per case | 1 | Nothing |
| Review | Nothing | none | Adjusting the budget |
Which controls keep spending safe?
Hard caps per agent. A spend limit at the provider or platform stops runaway use, such as an agent stuck repeating calls. Evidence: cap settings and alerts.
Cost per case, not total. Total spend rises with volume; cost per case shows whether the agent is getting more or less efficient. Evidence: the monthly report.
People's time counted. If held items take more time than they save, the budget shows it. Evidence: approver time per held case.
Spend through approved routes. AI subscriptions paid on cards bypass both budget and governance. Evidence: expense rules that route them to review.
Which signs show it is going wrong?
- Cost per case rises without a change in work. Prompts are growing, sources are being read in full each time, or the agent is looping.
- Held items cost more than they save. The agent works at a depth where people redo most of the work.
- Spikes at night or weekends. Something is running without anyone watching.
- Certification is cut to save money. The renewal is skipped or shrunk, and depth stays where it was without evidence.
- Cost moves between budgets. Agent costs sit in IT while savings are claimed in operations, so nobody sees the net.
What should you measure?
Cost per case by capability, held items per case and approver minutes per held item, model usage per case, spend against cap, and net saving per case compared with the work done without the agent. Ways to bring unit cost down are covered in how do you reduce the running cost of AI agents?
Where does it usually go wrong?
The pilot sets the price. A pilot on clean, simple cases suggests a low unit cost. Production includes the hard ones, which use more calls and more holds.
Savings are counted at full depth. The business case assumes the agent will finish cases on its own, while it starts at depth 2. Budget for the depth the agent will actually have.
Fixed costs are ignored. Certification, records and the operations team cost money even in a month with low volume.
No reserve. An incident that forces suspension sends work back to people at short notice. Plan for it.
Questions people ask
Should each agent have its own budget?
Each capability should have a measured unit cost, and each agent should have a cap. Budgets can be held by process, so an owner sees the whole cost of the work across agents and people.
How do we compare agent cost with staff cost fairly?
Compare cost per case done to the same standard, including review and rework on both sides. A case done cheaply by an agent and then corrected by a person costs both.
Does a higher depth always reduce cost?
Usually it reduces people's time on held items, but it raises the cost and stakes of mistakes. Move depth with evidence, not to meet a budget target.