A few people on your team have a Claude or ChatGPT seat, and someone always asks for another seat or a bigger API budget. You don't know if that's waste or a good investment, so you default to one flat rule: everyone gets the same cap, or nobody gets more. There's a way to check this person by person, instead of by blanket policy.
In this guide: a four question test that sorts any request by behavior you've already seen, and tells you what to do with it.
Start with the basics
The idea comes from finance, not AI: a credit card company doesn't give everyone the same limit on day one, and the limit isn't fixed forever. It rises for people who pay on time and spend it well, and it drops for people who miss. Cognition, the company behind the coding agent Devin, built exactly this: it turns every work session into an estimate of equivalent human hours, checked it against 126 users across 258 sessions, and found a 0.74 correlation with what users actually reported. Anyone who clears the bar gets a bigger budget automatically, without asking.
The budget check for anyone on your team
Why this works: a credit limit isn't fixed, it moves with real behavior. Once you look at output instead of the request itself, the decision gets a lot easier.
- Can you point to something concrete they produced with AI this month, that you'd otherwise have paid a person to do?If yes, that's a payment on time. The budget can go up.
- Do they flag where they're unsure about the output, before you have to ask?If yes but there's no big win yet, keep them at the current budget for another month.
- Has their AI use recently created extra work for someone else, like a mistake someone had to catch after them?If yes, that's a bounced payment. Freeze or cut it, no matter how good they looked before.
- Is this a first request, with no track record either way?Give them a standard budget, not the maximum, and check again in a month.
Try it on two real people on a team. A marketing coordinator used her Claude seat to produce three campaign drafts the client approved with no edits, that's question 1, and her budget goes up. A new sales rep got a seat two weeks ago and produced a proposal with a wrong price that a manager caught at the last minute, that's question 3, and his budget stays where it is or drops, no matter how hard he's trying.
This isn't theoretical. At Cognition, the system already runs this way, not as an idea. Russell Kaplan, the company's president, points at why this matters now: “per-person token spend is now approaching or passing that person's salary” at some companies. At that scale, an AI budget can't stay a fixed line item anymore.
Next ten minutes: pick one person on your team with an AI seat or budget, run the four questions, and decide now, instead of leaving it as it is because that's easier.