Spending limits
Seat limits and cost ceilings that enforce the budget, without ever measuring usage.
Your aim page set a spending ceiling: the most you'll spend on AI tools and builds this loop. A ceiling written on a page doesn't stop anything. This lesson turns it into limits the tools enforce, so nobody has to watch the bill every day, and nobody gets a surprise at the end of the month.
AI costs are easy to get wrong
Seats are predictable. Usage isn't. Many AI tools now charge by how much they're used, and an agent can use more in a night than a person does in a month. The numbers say most organisations are guessing.
the error organisations can make in cost estimates if they don't understand how generative AI costs scale.
Gartner, 2024
of IT leaders reported unexpected charges tied to consumption-based or AI pricing.
Zylo, SaaS Management Index, 2026
Uber had already used up its whole 2026 budget for one AI coding tool, as use blew past expectations.
Business Insider, May 2026
Two limits
The pilot kit sets two, and together they enforce the ceiling on your aim page.
- A seat limit per team. Each team has a number of paid seats. More seats come through the front door with a reason, like any other spend. It stops licences multiplying quietly, team by team.
- A monthly cost ceiling per tool, with an alert at 80%. Each tool gets its own cap. One overall budget hides a runaway inside an average. A cap per tool shows exactly which one is climbing.
The 80% alert is the circuit breaker. It gives the person who runs the loop time to look before anything stops: is this a tool stuck in a loop, or a team that found a genuinely valuable new use? A runaway gets switched off. A good new use gets raised at the front door, and the cap can move, inside the ceiling. If nothing is done, the hard cap stops the spend at 100%. That's why question 8 on the vendor checklist asks whether a tool has a hard cap at all.
Set the caps for your own tools below. They read the ceiling from your aim page and the tools from your approved list. Leave some room: the ceiling covers builds and new tools too.
Not set yet. Set it in module 1, or enter a monthly cap here.
Example tools. Add yours in lesson 1 and they replace these.
Set a monthly cap to see whether the tool caps fit inside it.
Setting the numbers
You won't get the caps right first time, and you don't need to. Start each tool's cap at its seats multiplied by the price, plus a margin for usage-based charges. Keep the total below the monthly share of your ceiling, because the ceiling also covers builds. The experiment budget is set separately on the aim page.
Send the 80% alert to someone who can decide the same day: usually the person who runs the loop, with finance copied. Then look at the caps every month for the first loop. A cap that is never close to its alert is probably too high. A cap that trips every month is telling you something: either a use worth paying for, which goes to the front door, or a habit worth fixing, which goes on the friction list.
Three tools, each with its cap. Around day 12, one of them starts running away: an agent stuck in a loop, or a team that found a heavy new use.
Each team gets a number of paid seats. More seats come through the front door with a reason, the same way any new spend does.
Usage is never a performance measure
The fourth principle of the framework is measure outcomes, never usage. Spending limits watch the money, per tool. They never watch people.
In 2026, several large technology companies learned what happens when usage becomes the target. Internal leaderboards ranked staff by how much AI they used, and people used more of it for its own sake (The Pragmatic Engineer, 2026). Amazon shut its leaderboard down in May. One of its senior vice presidents told staff: “Please don't use AI just for the sake of using AI. Use AI to help you solve customer problems, to help you solve business problems, to innovate” (Business Insider, 2026).
The control that works looks at spend, not people. Shopify's head of engineering described circuit breakers that cut off access when spending spikes within a day, to catch “runaway agents”. He added that they had found bugs in their own systems that way too (The Pragmatic Engineer, 2026).
The rule for your guardrails is short. Spending data goes to the person who runs the loop and to finance, as totals per tool. It never becomes a ranking of people, a target, or a line in anyone's review. What you measure is the goal on your aim page, at the bottleneck you'll find in module 4.
Most business AI plans let an admin set seat counts, and many let you set a monthly spending cap or usage alerts in the admin console or the billing settings. Where a tool can't cap itself, set the alert in your finance system and treat the 80% line as the point where someone decides. If a tool has no cap and no alert, that's a no on vendor question 8.
Adoption is slower than the board hoped. A director suggests publishing each person's AI usage every week, so the keenest users are celebrated and the rest are nudged to catch up.
What do you do?
Cap the money per tool, with an alert at 80%. Never measure the people.
One tool hits its 80% alert on day 18 of the month. What happens next?
Mark it complete. Your answers are already saved in this browser.