x
Because the moment something bills you per job rather than per month, it behaves like a supplier — and you already know how to manage suppliers. You don't manage a tool. You buy it, you use it, and if it's rubbish you cancel it. That's the correct amount of thought for a screwdriver, and it was roughly the correct amount for a flat-fee chatbot.
A supplier is different. A supplier does work on your behalf, invoices for it, and can affect your customers without you in the room. You give a supplier a scope. You agree a rate. You check the work. You review the relationship on a date. None of that is bureaucracy — it's the minimum that stops a working arrangement drifting somewhere you didn't choose.
An AI agent that answers customers, files documents or drafts client work now does all four of those things. The framing change is the point: the tool became a supplier when the pricing became per-job, and most people haven't updated the paperwork, because there wasn't any.
Four lines. One page. Resist making it five — the format survives because it's small enough to actually write.
Scope — these tasks only. Name what it handles and, more importantly, what it does not. "Answers questions about opening hours, delivery times and returns. Does not quote prices, agree deadlines, or discuss anything about an existing order."
Rate — the monthly ceiling. A real number, set in the tool, not an intention. This is the line that connects to the spend cap you set on Tuesday; a rate you haven't configured is a rate you've agreed to pay whatever it turns out to be.
Check — who actually reads the output. A named person and a frequency. "I read the overnight conversation log every morning with coffee." Not "we monitor it", which means nobody does.
Review — a date in the diary. Not "periodically". An actual date, because the alternative is that the arrangement renews by default forever.
Suppliers change their terms, and they don't ask permission. This is not a hypothetical: Meta's Business Agent moved to token billing at $2 per million tokens on 1 August 2026 after a free window that lasted thirty-one days, and from 1 October 2026 Meta resumes charging for service messages that have been free since November 2024.
Read that as a supplier story rather than a tech story and it's completely ordinary. Your supplier ran an introductory offer, ended it on schedule, and has notified you of a second price change taking effect in the autumn. Any business would diarise that. Almost nobody diarises it when the supplier is software, because software felt like a purchase rather than a relationship.
The review date is what catches it. Twice a year, per agent, ask four questions: is the scope still right, has the rate changed, is anyone still reading the output, and would I sign this again today? The last one is the honest one.
It's worth being clear about what the review is not. It isn't a performance check on the AI — you'd notice quickly enough if the output got worse. It's a check on the arrangement: whether the job you handed over is still the job being done, whether the price is still the price you agreed, and whether the oversight you promised yourself is still happening. Arrangements drift far more quietly than software does.
The contract exists to prevent these, and they're quiet precisely because nothing breaks.
Scope creep into client work nobody approved. It starts answering questions it wasn't given, competently and wrongly, and the first you hear is from a customer quoting it back at you. A written scope is what makes that reviewable — the same boundary logic as automating the task but never the promise.
A bill nobody owns. The charge is small enough to ignore for months, so it is ignored for months. Per-job pricing makes this more likely, not less, because each individual charge is genuinely trivial.
Outputs nobody reads until a customer does. The most expensive of the three. An automation that runs unattended is fine; an automation nobody has read since setup is a claim about your business that you haven't checked.
Keep it on the page, in plain words: drafts only, never decisions. No unreviewed contact with a client. You sign it, not the model.
That's the line that keeps the arrangement honest when it works well — and it works well often enough that the temptation to widen the scope is real. Write the contract before you're impressed by it. Ten minutes now, or an awkward email later; the cost difference is not close.
If you keep a decision log, the contract belongs beside it — one records what you decided to hand over, the other records why.
Not financial advice — vendor terms and rates change; verify current pricing and terms directly with each provider.
For the vetted tool list behind our own agent setup — free, 5-min read: 10 AI Tools That Save 10 Hours a Week.
Related Decodes
AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.
TL;DR: Once an AI agent bills per job rather than per month it is functioning as a supplier, so manage it like one with a one-page contract: scope (these tasks only, and explicitly not those), rate (a monthly ceiling actually configured in the tool), check (a named person and a frequency), and review (a real date). The review date is what catches supplier-side changes you didn't choose — Meta ended a 31-day free window on 1 August 2026 and resumes service-message charges on 1 October. Keep the human clause: drafts only, never decisions, and no unreviewed contact with a client.
