An internal knowledge assistant: what it really takes

An internal knowledge assistant answers staff questions from your own documents, with the source shown, and says "not in our documents" when nothing matches. What it takes from you is the set of documents that hold the real answers, a decision about who may see what, and a person who owns keeping the documents current. What it takes from the builder is retrieval that finds the right passages, answers restricted to those passages, and a test set of real questions. A two-week pilot on one team's top hundred questions shows whether it's worth extending.

What it is, in plain terms

Staff ask questions all day: "what's our cancellation policy?", "how do I process a return?", "what's the price for the large package?". The answers exist in procedures, price lists, old emails and someone's head. The assistant reads the documents you connect, finds the passages that match a question, and writes a short answer from those passages only, with a link to where it came from. It doesn't know anything you didn't give it, and that is the feature.

What it needs from you

1. The documents. Procedures, policies, price lists, product information, and past answers from chat or email. Formats don't matter much; accuracy and currency do. One outdated price list poisons every answer that touches it.
2. Access rules. Which documents are for everyone, and which for specific roles. The assistant must respect the same boundaries as your file system.
3. An owner. One person who updates documents when things change and reads the "not found" log monthly. Without an owner the assistant decays in three months.
4. Real questions for testing. Fifty questions staff actually asked, with the correct answer and where it lives.

How it stays accurate

• Answers come only from retrieved passages; no general knowledge.
• Every answer shows its source, so staff can check in one click.
• A clear fallback when nothing matches, with whom to ask.
• Documents are re-indexed when they change, automatically.
• The "not found" and "wrong" logs feed the monthly document update.

Where it fails

• Contradictory documents. Two versions of a policy; the assistant can't know which is current. Clean up first.
• Questions that need judgement. "Should I give this customer a discount?" is not in a document. The assistant should say so.
• Stale sources. The most common failure, and a documents problem, not an AI problem.
• Over-broad access. Connecting everything, including HR files, without rules.

A two-week pilot

After the pilot: extend to a second team, add role-based access, and consider the customer-facing version.

Costs to expect

The build sits in the middle-to-upper range of the cost guide, because retrieval and testing take real work; monthly costs are modest (AI usage, hosting) unless volume is large. The largest cost is usually the document clean-up, which pays off whether or not you keep the assistant.

Best fit and not a good fit

Best fit: businesses with ten or more staff, written procedures, and the same questions asked repeatedly: clinics, retail chains, agencies, logistics, professional services. Not a good fit: a team of three who sit together, or businesses whose knowledge isn't written anywhere yet; write it first.

What to do next

List the ten questions your team asked most last week and where each answer lives. If the answers exist in documents, find a builder on the marketplace who shows a retrieval assistant with visible citations, and ask for the pilot above.