Inside an AI context pack: what your AI needs to know about your business

I've written before about the AI context gap: ask a generic model for a business plan and you get someone else's, fluently worded. The fix isn't better prompting tricks or a more expensive subscription. It's a context pack: a structured brief that teaches the AI your business before you ask it anything. Here's what's actually in one.

The five layers

A working context pack has five layers. Facts: what you sell, to whom, at what prices, with what capacity and constraints. Economics: your margins by product or service, where money is made and quietly lost, the numbers you check weekly. Voice: how you actually write and speak, with real examples of your best work, plus a blunt list of phrases you'd never use. Customers: your real segments, their objections in their own words, lifted from actual emails and reviews rather than imagined. And boundaries: what the AI must never do, what data never goes in, and the brand and quality rules that keep slop away from anything a customer sees.

What it changes

The difference is immediate and slightly eerie. Ask an uncontexted model to draft a follow-up to a lapsed customer and you get beige: "I hope this email finds you well." Give it the pack and it knows the customer bought twice last spring, that your tone is dry and direct, that you never discount but do prioritise repeat customers for scarce slots, and the draft comes back sounding like you on a good day. Analysis improves even more than writing: "which of these three jobs should we take?" becomes answerable when the model knows your margins and your capacity.

Build a starter version yourself

You can get 60% of the value in an afternoon. One page of facts and numbers. Three or four examples of your best writing, pasted whole. Your ten most common customer questions, with the answers you actually give. Put it in a single document, and give it to your AI tool at the start of every session; most tools now have a projects or instructions feature that stores it permanently. Update it monthly, because a stale pack fails quietly.

Where the last 40% lives

The hard part isn't the document; it's the judgement around it. Which use cases genuinely move your numbers, rather than merely feeling clever. Guardrails that survive contact with a busy Tuesday. Wiring the pack into the tools your team already uses, so it gets used rather than admired. And getting the team to trust it, which is a change problem, not a technical one. That's what the AI Context Sprint does: two weeks, fixed price, and you keep everything, including the pack itself.

Ian Sharpe

Fractional executive and portfolio consultant. The person you meet is the person who does the work. Who am I? →