Where is the spend cap actually hidden?

Usage-based AI pricing has no brakes fitted as standard. Here's where the spend cap hides in every tool, what number to set, and how to pick it honestly.

In almost every metered AI tool it lives three clicks deep, in the same place: Settings → Billing → Limits (sometimes "Usage limits", "Budget" or "Spend controls"). You set a hard monthly cap, which stops the service when it's hit, and a soft alert partway up, which emails you before it does. Both matter. The alert is what gives you a decision; the hard cap is what saves you when you're not reading email.

Ten minutes, once, across every tool on your card. That's the whole job. The reason almost nobody does it is that the cap is not on the pricing page, not in onboarding, and never suggested at the moment you switch an automation on — the moment it would actually help.

Why does this matter more than it did last month?

Because the pricing model changed underneath you. Meta's Business Agent moved to token billing at $2 per million tokens on 1 August 2026, roughly 4–5 cents per conversation. OpenAI moved workspace-agent and spreadsheet tasks onto token-based credits in early July. Microsoft put 365 prices up on the 1st.

Note the shape of Meta's rollout in particular: the free window opened on 1 July and closed on 1 August — thirty-one days. That's the pattern worth internalising. Free windows close on a date somebody else picked, and the switch from free to metered does not require you to agree to anything. Your automation keeps running. It simply starts costing.

A flat subscription failed safe. If you overused it, you got throttled or nagged. A meter fails expensive: it does exactly what you asked, more times than you expected, and tells you afterwards. The brakes are not fitted as standard. You fit them.

What number should the ceiling be?

Not "a bit more than last month". Set it from the value, not the usage:

Hours the tool saves you per month × your hourly rate. Then halve it for month one.

If a tool saves four hours a month and you charge £60 an hour, it's producing £240 of value. A £120 ceiling in month one is generous and still leaves the tool comfortably profitable. Raise it later on evidence — because you hit the cap doing work that paid — rather than on optimism, which is how ceilings quietly become decoration.

If you can't state the hours a tool saves you, that's the finding. You've just discovered a tool you're paying for on vibes, and no cap will fix that. Score it properly first: minutes back per pound is one number, and a tool that can't produce it is a cancellation candidate, not a capping candidate.

Set the alert at roughly half the hard cap. Half gives you a fortnight of warning at a steady burn rate — enough time to look at why usage climbed, which is usually more interesting than the bill. Usage doubling is either a genuinely busy month, or something looping.

What happens when you hit the ceiling?

This is the part to check before you need it, and it differs by tool. Three behaviours are common, and only one is safe:

• Hard stop. The service refuses further calls. Safe, and occasionally embarrassing — an autoresponder that goes quiet mid-conversation.
• Alert only. The cap is advisory. It emails you and keeps spending. This is not a cap; treat it as an alert and manage the risk elsewhere.
• Downgrade. Falls back to a cheaper model or a queue. Usually the best outcome, if offered.

Find out which one you've got, and write it down next to the cap. Where a customer-facing automation would hard-stop, decide now what it should do instead — an out-of-hours message, a form, a human handover. An automation that fails silently in front of a customer is worse than one that costs $12 more, and that trade-off is a decision you should make deliberately rather than discover on a Saturday. It belongs in the same place as your escalation rule: the written line where the machine stops and you start.

The ten-minute version

Open your card statement. For every AI tool on it: Settings → Billing → Limits → set the hard cap, set the alert at half, note what happens at the ceiling. Then set a calendar reminder for the first week of next month to read the actual spend against the cap.

That last step is the one that turns a cap into a control. A ceiling nobody reviews just tells you whether you had a catastrophe; a ceiling you compare against actual spend each month tells you whether the tool is growing into its budget for good reasons or bad ones. Two months of data is usually enough to know which, and it takes about ninety seconds to check.

Do it before you switch the next automation on, not after. A ceiling set in advance is a budget. A ceiling set after a surprise invoice is a lesson, and it's a more expensive way to learn the same thing. The news that pushed this up the list took effect ten days ago.

Not financial advice — vendor rates and cap behaviour change; verify in your own billing dashboard.

For the prompt set we use to run this review each month — free, 5-min read: 7 AI Prompts That Run a One-Person Business.

Related Decodes

• AI Just Became a Utility Bill — the week's hub: the 5-minute bill audit this cap sits inside.
• Your Chatbot Now Has a Meter — the pricing change that made ceilings urgent.
• Safe to Automate — deciding what should run unattended in the first place.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

TL;DR: Every metered AI tool hides a spend cap under Settings → Billing → Limits — set a hard monthly ceiling plus an alert at half of it, in every tool on your card. Pick the number from value rather than usage: hours saved per month × your hourly rate, halved for month one, raised later only on evidence. Check what each tool does when the cap is reached, because "hard stop", "alert only" and "downgrade" are very different outcomes for a customer-facing automation, and only one of them actually stops the spend.