Build the AI backend once. Bill it forever.

The problem: you keep building the same thing

Client one wants a support bot. Client two wants a product-recommendation assistant. Client three wants "something like ChatGPT but for our docs." Three different products — and underneath, the identical plumbing: store conversations, keep the API key out of the client bundle, track what it costs, stop one enthusiastic end-user from burning the budget, and answer the client's inevitable question: "what is this AI thing costing us?"

Most agencies rebuild that plumbing per project. It's a week or two of un-billable-feeling infrastructure work every time, it's slightly different every time (so every maintenance ticket is a re-learning exercise), and the cost-tracking part usually gets skipped — which comes back to bite at invoice time.

The shift: every client is a project

Cerver's model maps one-to-one onto agency life. Each client gets a project: its own API keys, its own vault for the provider key, its own budget, its own dashboard section. Your integration is the same two calls on every engagement:

# open a chat for one of the client's end-users
POST /v2/sessions        { "app_user_id": "their_user_42" }

# send a message — reply streams back, with its cost
POST /v2/sessions/{id}/run-llm   { "input": message }

That's the whole thing. The developer who wired client one's support bot already knows how to build client five's document assistant — the products differ, the platform doesn't. (Watch this live on the platform page — two products running through the same two calls, and you can add a third yourself.)

The money part, concretely

Two things quietly decide whether AI features make or lose you money:

1. You can only bill what you can attribute. When the model spend for all clients runs through one shared OpenAI or Anthropic account, the monthly invoice is one number — and carving it up per client is a spreadsheet guess. Guesses get rounded down, absorbed, or disputed. With a project per client, the number is exact and queryable: Client A's assistant cost $214.60 this month, across 1,840 conversations. That's a pass-through line on an invoice, plus your margin on top if that's your model.

2. One end-user can eat a retainer. Every provider cap is account-wide — it protects the provider's bill, not any single client's budget. A per-customer cap at the platform level means the 501st over-budget message is refused before it's charged. Your fixed-fee project stays fixed-fee.

The handoff gets better too

Agencies live and die on clean handoffs. Because each client is a self-contained project, handing over is transferring ownership of that project — the client gets the dashboard, the conversation history, the cost view, and their own provider key in their own vault. No untangling their data from your account, no "we'll export the logs." And if you keep the retainer, you manage all clients from one place: one dashboard, every project, cost per client at a glance.

What to do this week

Learn it once, on something small: the quickstart is two calls and takes minutes. Then, on the next brief with "and an AI assistant" in it, open a project named after the client instead of opening a blank backend/ folder. The second time you do it, you'll notice the estimate you're writing is a day, not a fortnight — price accordingly, and keep the difference.