Building a Knowledge Base Your AI Can Actually Use

The single biggest determinant of whether a customer-facing AI is useful is what it has to read. Our guide to teaching an AI to sound like your business puts the knowledge base first for that reason: rules and tone can be adjusted in an afternoon, but an AI with nothing true to retrieve will be wrong in ways no amount of prompting fixes.

How retrieval actually works

When a customer asks something, the system does not send your entire document library to the model. It finds the passages that look most related to the question and puts those in front of the model as context. Everything about how you write your documents should serve that search.

This is why a forty-page PDF brochure performs badly. The relevant sentence about cancellation fees is buried among design language and photography, so the passage that gets retrieved is half marketing copy and half the actual answer. Short documents on single topics retrieve cleanly; long documents on many topics retrieve badly.

Start with the questions you already get

Do not begin by writing documents. Begin by reading your inbox. Pull the last two hundred customer conversations and list the questions that come up more than twice — you will usually find that fifteen to twenty questions cover the large majority of everything ever asked. Those are your first twenty documents.

Write each one with the customer's phrasing in the title. "Can I bring someone with me?" retrieves better than "Guest policy", because the customer's question and your title use the same words. This is unglamorous and it is most of the work.

What belongs in there

• Prices, including what is and is not included, and how you handle quotes for anything variable.
• Policies with the actual numbers — cancellation windows, deposit amounts, refund conditions, warranty length.
• Practicalities: opening hours, parking, access, what to bring, how long things take.
• Service descriptions in plain terms, including who each one is not suitable for.
• The things you say no to, written out. "We do not offer same-day appointments" prevents more damage than any positive statement.

“

Write the answers to the questions you actually get, in the words your customers use to ask them. Everything else is decoration.

— be digital ai team

What to keep out

The knowledge base is quotable to customers, so treat every document as something the AI might read aloud. Cost prices, supplier terms, margin notes and internal staff instructions do not belong in it. Neither does anything about individual customers — that is contact record territory, with its own retention rules.

Be careful with drafts and superseded versions. If last year's price list is still in the folder, retrieval will find it, and the AI has no way to know which of two contradictory documents is current. One document per topic, no archives.

Structure beats volume

Teams often assume more documents means better answers. Past a point the opposite happens: near-duplicate documents compete with each other and the retrieved context becomes a muddle. Thirty clean, distinct documents outperform three hundred overlapping ones, and they are dramatically easier to keep accurate.

Keep each document to a single topic and a few hundred words. If a document needs subheadings, it is probably two documents. And write in complete statements rather than fragments — "Deposits are 30% of the service price and are refundable up to 48 hours before the appointment" is retrievable and quotable; a table cell reading "30% / 48h" is neither.

Keep it current

A knowledge base decays quietly. Prices change, a service is discontinued, a policy is relaxed, and nobody remembers there is a document saying otherwise until a customer holds you to it. Give one person ownership, review it whenever a price or policy changes, and diary a full read-through twice a year.

The fastest way to find the gaps is to look at where the AI struggled. Every conversation it handed to a human is a question your documents did not answer — reading a batch of them, as described in reading the AI audit log, turns your knowledge base into something that improves rather than drifts.

Test before you trust

Once the documents are in, ask the AI the twenty questions you started from and check every answer against reality. You are looking for two failures: answers that are wrong, and answers that are right but sound nothing like you. The first is a document problem; the second is covered in writing system rules and a style guide for your AI.