The question every AI project hits in month two
The build is signed off, the feature is live, and the first provider invoice arrives. Somebody has to pay it, and unless the contract said exactly how, it becomes a conversation. There are three workable setups. The right one depends on how much usage risk you want and how much admin you can carry.
Setup 1 — The client owns the provider account
The client opens their own OpenAI, Anthropic or Google account, adds their own card, and gives you an API key for their project. The provider bills them directly.
- Good for: larger clients, regulated clients, and anyone who wants the usage on their own books.
- You earn on: the build and an ongoing support or optimization fee — not on usage.
- Watch out for: you lose sight of the spend unless every call goes through something you can see. Keep a per-client view of requests and cost anyway; when the client asks “why did our bill double?”, you want the answer before they do.
Setup 2 — Your account, rebilled with a markup
You run all clients on your own provider account and invoice each client for their share, usually at cost plus a margin.
- Good for: small and mid-size clients who don’t want another vendor relationship.
- You earn on: the markup, and it grows with the client’s usage.
- What it needs: per-client metering you can defend line by line — separate projects and keys per client, priced from the provider’s own token counts — and a per-client cap, because their overrun is now on your card first.
Setup 3 — Bundled into a fixed fee
Usage is included in a monthly retainer up to a ceiling. Simple for the client, risky for you if the ceiling isn’t enforced. See how to price AI work for clients for sizing the retainer from token math.
Contract clauses worth writing down
- Who owns the provider account, and what happens to keys and data when the engagement ends.
- The monthly ceiling and what happens at it: the feature pauses, or overage is billed at an agreed rate.
- What counts as usage: model calls only, or also retries, evaluations and test traffic.
- How usage is evidenced: a monthly report per client with requests, spend and model mix.
- Who benefits from optimization: if you cut the bill by caching or routing to cheaper models, whether the client keeps all of it or you share it.
Whatever the setup: meter, cap and report per client
All three setups fail the same way — nobody can say what one client’s AI cost last month, or stop it from doubling. Three habits prevent that: give every client its own project and key so each request is attributable; put a hard monthly ceiling per client that refuses requests before the provider is called; and send a monthly report the client can read. What to put in that report.
How ModelGate handles it
Each project in ModelGate holds its own provider keys, so a client’s project can run on the client’s account (setup 1) or yours (setups 2 and 3). Projects group under clients, with a spend cap per client enforced at the gateway, and a monthly report per client to back up your invoice. For ModelGate’s own fee you choose per client too: it goes on your invoice, itemized by client, or the client approves their own PayPal subscription from a billing link and is invoiced directly — you never see their card. ModelGate for agencies.
Keep reading
- How to price AI work for clients when model costs move every month — Retainer, cost-plus or a share of savings — and the token math that decides your margin.
- What to put in a monthly AI cost report for a client — The seven numbers a client actually reads, and how to get them without a spreadsheet.
- ModelGate for agencies — per-client ledgers, caps, reports and billing.