Stockertim Dictionary · Artificial intelligence · Business
Spend cap
Also called spending cap, spend limit, budget cap
In short
A spend cap is a hard limit on how much an automated system may spend on model use or paid actions per run, day or month, enforced outside the model so it holds whatever the model does.
- Beginner
- Concept
- Checked
In detail
Automations that call AI models pay for each use, usually by the amount of text sent in and generated, and some also pay per action, such as each text message or data lookup. Normally the cost per run is small. The danger lies in the unusual run: an agent stuck in a loop, a sudden flood of inputs, a misconfigured trigger firing thousands of times, or an attacker deliberately feeding long inputs to run up costs.
A spend cap bounds that danger. Caps are typically set at several levels: per run, so one item cannot consume much; per day, so a fault is contained within hours; and per month, matching the budget. When a cap is reached, runs stop or queue and a named person is alerted. Caps are set from measured costs with headroom, for example a few times the typical cost per run, so that legitimate unusual cases still finish.
The cap must sit where the model cannot override it: in the workflow tool's settings, the account with the model provider or the code that makes the calls. An instruction in the prompt to keep costs down does not limit anything.
Caps work alongside monitoring. Hitting a cap is a signal worth investigating, and a cap set so high it is never approached offers little protection.
Worked example
Hollybrook Insurance Brokers runs a document-summary agent that costs about £0.05 per run.
- Per-run cap at four times the usual cost: 4 x £0.05 = £0.20
- Expected volume of 300 runs a day: 300 x £0.05 = £15
- Daily cap with headroom: £30
- Runs the daily cap allows at the per-run cap: £30 / £0.20 = 150
A looping run is stopped at 20p, and a fault cannot cost more than £30 in a day before it is noticed.
In practice
Teams set spend caps per run, per day and per month in the platform or code that makes the model calls, alert an owner when any cap is reached, and revisit the caps when prices or volumes change.
Watch out
- A cap in the prompt is not a cap; it must be enforced by the surrounding system.
- A daily cap set too tight can halt legitimate work on a busy day, so it needs an owner who can raise it.
Learn it properly
Official sources
- OWASP Gen AI Security Project: LLM Top 10 · OWASP Foundation
- AI Playbook for the UK Government · UK Government
Facts last checked on 9 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.