A spend limit is a hard stop on what an agent may consume or commit before it has to hold and ask a person. Set it in money, at three levels: per run, per day and per decision type. Enforce it inside the system, where a run cannot get past it.
Most teams meter tokens, because that is the unit the vendor bills in. It is the wrong unit. Tokens measure what the model consumed. A limit has to measure what the business is exposed to. That includes everything the agent approved, refunded, ordered or paid.
Two limits, one for each exposure
Compute spend is what the agent costs to think. Commitment spend is what the agent costs you by deciding. They fail at different speeds, and a single limit cannot cover both. A runaway retry loop burns compute in minutes. A badly set approval rule commits money for two weeks before anyone looks.
| Exposure | Compute spend | Commitment spend |
|---|---|---|
| What it is | Tokens, calls, retries | Money the agent approves or commits |
| Typical scale | Small amounts per day | Large amounts, once it is wrong |
| Fails in | Minutes, loudly | Weeks, without noise |
| Caught by | A usage alert | A month end check, much later |
| Right limit | Per run and per day | Per decision and per supplier or customer |
| At the limit | Stop and page an engineer | Hold and ask a named approver |
The second column hurts most. It does not appear on the model provider's invoice, so cost tools often miss it.
GatehouseIn RUN-7F2K4, the simulated run on Gatehouse, Surehand's control plane, the manifest sets an autopay limit of 250.00. Rule bank-detail-change holds INV-40218 for 2,400.00 for the AP lead by name. Read the run on Gatehouse
How to set the number
Do not guess a budget. Work it out from the process the agent replaces, whose cost and error rate you already know.
- /01
Measure the current unit cost. What does one claim, invoice or ticket cost to handle today, including all staff time? Call this C.
- /02
Measure the current error cost. What does a wrong one cost, on average, including rework and goodwill? Call this E.
- /03
Set the per run compute limit at a small share of C. If thinking about the work costs a meaningful fraction of doing the work, the economics are already wrong.
- /04
Set the per decision commitment limit below E. An agent should never be able to commit more, unattended, than one mistake costs to clean up.
- /05
Set the daily limit a little above expected volume. That leaves room for a busy Monday, and a loop trips it quickly rather than overnight.
- /06
Review all three once the deployment has run on real inputs for a few weeks, when you have real numbers instead of estimates.
How to set AI spending limits by team
Give each team its own daily limit, owned by the person who answers for that budget. Finance sets the ceiling for accounts payable. The support lead sets it for refunds. Sales ops sets it for quotes. One company-wide number hides which team is burning money and gives nobody a reason to tighten it.
A team limit sits on top of the per run and per decision limits. It does not replace them. When a team hits its daily number, every agent working for that team holds until its owner raises the limit or the day resets.
Spend limits and approval thresholds are different numbers
A spend limit is the most an agent may commit at all. An approval threshold is the point above which it must ask first. An accounts payable agent might pay invoices under 250.00 on its own, hold anything from 250.00 to 10,000.00 for the AP lead, and refuse anything larger outright. That is one threshold and one limit, and both belong in the same policy.
Can an AI agent approve a discount on its own?
Only below a number someone with pricing authority wrote down. A quoting agent can apply the discounts your price list already allows. Anything above that band holds for the sales manager by name, with the quote, the customer and the margin attached. The agent should not be able to change the band itself. A discount is commitment spend: it costs nothing on the model invoice and real money on every order that follows.
What happens at the limit
At the limit the run should hold. A failure throws the work away and reports an error where nobody looks. The item goes back to a person who never knew it was tried. A hold keeps the run and records why it stopped. It goes to a named approver with the evidence gathered.
That is also what makes the limit safe to set tight. When hitting the limit is cheap and recoverable, you can afford to start conservative and relax it with evidence. When hitting the limit means lost work and an angry customer, everyone sets it high, which is the same as not having one.
- Compute, per run
- set from the unit cost
- Compute, per day
- expected volume, plus room
- Commitment, per decision
- below the cost of one mistake
- Commitment, per day
- agreed with finance
- At the limit
- Hold, notify the named approver
Four limits cover one process, and two of them are on money the agent moves.
A limit has to stop the run
A dashboard reports what already happened. A limit prevents it. If only a person watching a chart stands between a retry loop and your card, the control is their attention. Nobody is watching at 2am on a Sunday. The check has to run before each call.
Ask whether the system can exceed the number at all. If the honest answer is that it would be noticed quickly, that is monitoring. A limit means the run stops.
What a limit does not protect you from
A limit does not stop an agent from being wrong at low cost. An agent that sorts a thousand claims wrongly under budget has cost you nothing in spend and a great deal in trust. Limits bound financial exposure. They say nothing about accuracy. A team that watches only spend will miss a quality drift that stays inside every limit. Confidence thresholds and sealed records cover that, which is why the three are set together.
Common questions
How should merchants set spend limits and approval thresholds for AI agents?
Use two numbers per agent. An approval threshold: below it the agent acts alone, above it the action holds for a named person. A spend limit: the most the agent may commit at all, above which it refuses. Derive both from the process you are replacing: keep the threshold under the cost of one mistake, and set daily limits a little above normal volume.
What is the best way to set spend limits across different AI agents and tools?
Enforce the limit in one place that every agent passes through, not inside each tool. Limits set per vendor console or per CLI drift apart and miss money moved through other paths. A single check before each action, with per run, per day and per team numbers in money, covers every agent the same way and leaves one record.
What is an AI agent purchase limit?
The largest purchase an agent may make on its own. Above it, the order holds for a named approver with the supplier, amount and reason attached, and the agent cannot raise the number itself. Some payment providers now offer agent purchase limits too; treat theirs as a backstop to your own rule, not a replacement.
Read next
What is an AI audit trail?
Seven fields. And why logs don't count. What your risk team will ask for, and what most systems can't produce.
How to monitor AI agents in production
Uptime tells you it's running. Not that it's right. Four signals to watch: quality drift, hold rate, spend per run, latency.
