AI knowledge pack: the material your assistant should actually be grounded in

A knowledge pack is the company's answers made explicit: owned, versioned, and reviewed on a date rather than on a hunch.

Every assistant answers from something, and in most companies that something is whatever happened to be pasted in last week: a brochure here, an old email there, a pricing note someone edited twice and forgot. The answers come out confident regardless, because that is what these systems do. The honest alternative is a knowledge pack: one owned, versioned body of material the assistant is grounded in, built and maintained like any other company asset.

The first layer is the questions customers actually ask, paired with the answers the company has approved. Not the answers someone improvises at midnight, but the ones an owner has read and agreed to stand behind. Around twenty good entries cover the weekly traffic of most small businesses, and the long tail can grow after that. Writing them down is also the first time many companies discover the whole team has been answering the same question five different ways.

The second layer is the rules the assistant must not break: what it may quote and what it must never invent, which topics route to a person, and how it speaks. Tone lives here too, written down so it survives staff changes and vendor changes alike. An assistant without rules is not flexible; it is unsupervised.

The third layer separates the facts that change from the facts that rarely change. Offers, timelines, and policies belong in a section designed to be edited often; company background and definitions belong in one that can stand for months. Alongside both sit the escalation paths with named owners, so the assistant knows who receives what, and the team knows who decided it.

What makes it a pack rather than a pile is governance, and it is three small things. An owner: one person responsible for the material, whose name appears on it. A version: so anyone can say which text produced which answer, which is how you fix an answer instead of arguing about it. A review rhythm: a date on the calendar, not a feeling that things are probably current.

The honest framing is that the pack is never finished, and that is fine: it is the company's answers made explicit, and most companies have never done that for their own humans either, so building it pays twice. The failures come from skipping the governance. The assistant quotes an expired offer because the pack said so, in polished prose, all day long. Three documents inside the same pack contradict each other, and the assistant picks whichever it retrieved first. The one person who knew what the pack contained leaves the company, and the material becomes an orphan nobody dares to edit.

The practice is small enough to start this week: write down the questions that arrive weekly, draft the approved answers once, mark what must never be said without a human, give the pack an owner and a version number, and put the review on a date rather than on a hunch. Where Bloom sits is beside you in this work: we build assistants grounded in material the client owns and can read, not in a mystery box. Guidance, not guarantees: a pack makes an assistant accountable, not infallible, and any answer with money or legal weight still deserves a person.