← Back to Blog
How Do I Teach My AI Agent Each Client's Contract Terms, SLAs, and Scope Boundaries?

How Do I Teach My AI Agent Each Client's Contract Terms, SLAs, and Scope Boundaries?

November 13, 2026·6 min read

Your agent gives away work because it has never read the contract, or it read it once inside a chat that closed three weeks ago. The fix is not a bigger model. It is a per-client context file that loads before every conversation and says, in plain language, what is in scope, what is out, what the SLA clock is doing, and what to do when a request lands outside the lines.

I've watched this happen across my own companies. The agent isn't malicious. It's helpful the way a new hire is helpful on day two, before anyone tells them what the retainer actually covers.

Why does my AI agent keep over-promising on client work?

Because helpfulness without constraints is scope creep with good manners. The model was trained to say yes. It has no idea that the weekly report covers two dashboards and not three, so when a client asks "can you also pull the Shopify numbers in there," it says sure, happy to. You just built a data pipeline for free.

Do that four times across a year and it's real money. On a $4,000 a month retainer I've seen the effective rate drop by a third from small yeses nobody logged.

The second cause is boring. The contract lives in a PDF in a shared drive that no agent has ever indexed. Even if it did, contracts are written for courts, not for models. The agent needs the operational version: quantities, cadence, exclusions, response clocks.

What belongs in each client's context file?

Seven blocks per client. Keep the whole thing under about 600 words so it actually fits in context.

Scope in. Verbatim deliverables with quantities and cadence. "Two reports per month, delivered by the 5th, covering paid search and organic."

Scope out. The block most people skip. "No creative production. No landing page builds. No Shopify or ecommerce data. No weekend coverage." This is where the agent earns its keep.

SLA table. First response, resolution target, business hours, and what pauses the clock. Write it as a table, not a paragraph.

Rate card. Out-of-scope rates, so the agent can quote instead of guess. Never let it invent a number.

Approvers. Who can authorize new work and who cannot. Half your scope creep comes from a client's junior employee with good intentions.

Relationship notes. Who they are, what they care about, what went wrong last quarter, whether they want a call or a written summary.

Escalation rule. One sentence the agent uses when a request isn't covered: "Let me check with Joe and come back to you today."

I run this as a Notion database with one page per client, mirrored into a Google Drive folder as plain markdown. Markdown beats a PDF every time. Airtable works too if you like structured fields and review views.

How do I make sure the right context loads every time?

This is context engineering for founders, and it's mostly architecture, not prompting.

Do not build one giant prompt that holds every client. Build a router. The agent identifies the client first, from an email domain, a form field, or a project ID, then loads only that client's folder. In Claude, that's one Project per client with the context file in project knowledge. In ChatGPT, it's a custom GPT per client for your top accounts. If you're running agentic workflows for service businesses at any volume, n8n or Zapier does the routing: webhook in, client ID out, context file attached, then the model call.

What's becoming clearer to me is that people overbuild the retrieval layer and underbuild the boundary layer. You do not need a vector database for 18 clients. Pinecone or Supabase pgvector earns its place at a few hundred clients or thousands of documents. Below that, folders and a router will beat it, and you'll be able to read exactly what the agent is seeing.

That's the whole point of an AI agent operating system. The context lives somewhere you can inspect, version, and correct. The agent is thin on top of it.

How do I test whether it will hold the line?

Red-team it before a client ever touches it. Write 20 prompts per client that push on the boundary, run them, and log the results.

Use the soft asks, the ones that sound like favors. "Quick one while you're in there." "Can you just add one more chart." "I know it's not in scope but could you." Your agent will catch the obvious violations. The soft asks are where it folds.

Log pass or fail in Airtable or a Google Sheet, one row per prompt, with the exact response. Run the suite in Claude or ChatGPT against the real context file, not a summary of it. If you want to script it, Cursor and a 40 line Python script will run the whole suite in a minute and diff one version against the next.

The check that matters most: does the agent commit, or does it acknowledge and escalate? "Sure, I'll add that" is a fail. "Let me confirm with Joe" is a pass.

How do I keep the context current when a contract changes?

Every amendment goes into the file the same day. Otherwise the agent keeps operating on an old truth, quietly, and the first person to notice is the client.

Put a version line at the top of every file: "v4, updated March 12." Then the agent can state its own freshness, and you can see at a glance which clients are stale.

Give one person a 30 minute pass on the first Monday of the month. An Airtable automation with a review date field, or a Zapier reminder into Slack, makes it harder to forget. Check renewal dates, expired promos, new deliverables, and every "we'll figure it out later" that finally got figured out.

I've been building businesses since 1995, and the pattern is always the same. The things that scale are the things you wrote down. The people who scale are the ones who go back and review what they wrote down.

FAQ

Can I just paste the contract into the chat instead of building a file? You can, and it half-works. The chat resets, the model misreads legal language, and you'll re-paste it forever. Extract the operational terms once into a context file and the contract becomes an attachment you rarely touch.

Should the agent quote prices for out-of-scope work? Yes, but only from your rate card, and only as a range. Anything unusual gets escalated. An agent that invents a number is worse than an agent that says nothing.

How many clients can one agent handle this way? The folder and router approach holds up comfortably to about 40 clients. Past that you need real routing logic and retrieval, and you'll want to split by service line before you split by tool.

If you want the exact structure, I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book.

Build your AI operating system from the book

Get Joe Che's AI OS book and turn these ideas into a practical operating system for your work.

Prefer to build it live with Joe? Join the AI Business Mastermind