Once people accept that their AI needs a brain, the next mistake is nearly always the same one: they try to feed it everything. Every email, every document, every note, on the theory that more information must mean smarter answers. It does not. A knowledge base is not a loft you cram full and forget. It is a working reference, and a working reference is only as good as it is clean. This guide is about the two decisions that actually matter: what earns its place, and what quietly makes your AI worse.

If you have not built one yet, our guide on building a second brain this week is the practical starting plan. This one is about what goes inside it.

The rule underneath everything: specific, stable, true

Before the lists, here is the single idea to hold on to. Good context is specific, stable and true across many jobs.

Specific, because vague inputs give vague outputs. “We offer competitive prices” teaches the AI nothing. “Bathrooms start around a figure and I quote once I have two photos” teaches it plenty.

Stable, because the AI leans on this constantly, so it needs facts that hold across lots of jobs rather than one-off details that were true for a single customer last Tuesday.

True, because a confident wrong answer is worse than no answer, and the AI will state whatever you give it as gospel. Everything below is really just this rule applied.

What earns its place

These are the things that pay their way, because the AI reaches for them again and again.

  • Your prices and how you scope work. Starting figures, packages, day rates, the rule you use to price a job. This is the single most valuable section, because it is what the AI invents most badly when it is guessing.
  • Your policies. Deposits, payment terms, lead times, what you do and do not take on, how you handle complaints and changes. The load-bearing facts of how you run.
  • Your voice. A few real messages you have sent, so drafts come out sounding like you. Examples work far better than describing your tone in the abstract.
  • Your standard answers. The questions customers ask every week, with the answers you actually give. These become the raw material for replies and FAQs.
  • Your products or services in detail. What each one is, who it suits, what it includes, common objections and how you answer them. The things a good salesperson would know cold.

Notice the pattern. Every one of these is stable, gets used constantly, and is something the AI would otherwise make up. That is what belongs.

What to leave out

This is the half people skip, and it is the half that protects the quality of your answers.

  • One-off details. The specifics of a single job that will never recur. Interesting to you, noise to the AI. They bury the facts that matter under facts that do not.
  • Anything out of date. Old price lists, retired services, superseded policies. Stale facts are not neutral, they are landmines, because the AI cannot tell they expired and will quote them with total confidence.
  • Contradictions. If two documents say different things, the AI has no way to know which one wins, and its answers will wobble between them. Reconcile them before either goes in.
  • Half-finished thinking. Draft ideas, maybes, notes to self. If it is not settled, it will make the AI hedge and waffle. Put in what is decided, not what you are still chewing on.
  • Sensitive data you do not need. Customer personal details, card numbers, anything private that is not required for the AI to do its job. If it does not need it to answer well, keeping it out is simply safer. When in doubt, our wider material on data safety is worth a read before you paste.

We learned this the slightly hard way running a persistent context system of our own. Early on the temptation was to pour everything in. We got far more value from a smaller, cleaner brain than a bigger, messier one, because the AI could actually tell what mattered. The candid version of that lesson is in how we run our own AI brain.

The one-line test

When you are unsure about a document, ask one question: will this fact be true and useful across lots of jobs?

If yes, it belongs. Your deposit policy passes. Your standard bathroom pricing passes. Your tone examples pass.

If no, leave it out. A single customer’s quirk fails. Last year’s prices fail. A half-formed idea fails. Anything sensitive you do not strictly need fails.

That one question sorts the vast majority of decisions in seconds, and it keeps your knowledge base doing the job it exists for: giving the AI a clean, current, trustworthy picture of your business to answer from.

Keep it lean, keep it current

A knowledge base is not finished when you build it, it is tended. Every time something changes in the business, the fix is to update the source and the answers update with it. No retraining, no rebuild. And every so often it is worth a quick read-through to prune anything that has drifted out of date. Lean and current beats big and stale, every time.

Where to take this next

Our flagship course Build an AI-ready business: give your AI a brain has a full lesson on deciding what goes in and what stays out, with worked examples for real trades.

To get started on the right footing, the AI Readiness Checklist on our free resources shelf helps you pick the first things worth writing down.

And if you would rather sort your knowledge base in a room with people who build these systems for a living, come to our AI Automation Masterclass in Manchester. Tickets are normally £20. This one’s free, a limited-time offer to launch the series.