Building a knowledge base your AI assistant can use

Why the knowledge base is the whole ballgame

An AI assistant that answers customer texts is only as good as what you have told it. Give it accurate, current information about your business and it answers like a well-briefed staff member. Give it nothing, or worse, out-of-date scraps, and it either says "I'm not sure" to everything or, if it is a poorly built one, makes something up. For a business, a confidently wrong answer is far more dangerous than an honest "let me check".

This is what grounding means. A grounded assistant answers from your own knowledge base rather than from general internet guesses. It says what is true about your business and stops there. That is the behaviour you want, and it is entirely down to the work you put into the knowledge base. The good news is this is not technical work. It is writing down what you already know.

Start with the questions you already answer

The fastest way to build a useful knowledge base is to work backwards from real questions. You and your team already answer the same things every day. Those answers are your first entries.

Sit down and list the questions customers actually ask. For most small businesses the core set looks like this:

• What are your hours, including public holidays and any seasonal changes
• Where are you, and is there parking or an entrance that is easy to miss
• What do you charge, or how does your pricing work, including call-out fees
• What services do you offer, and what do you not do
• How do people book, pay, and cancel
• Your policies: deposits, cancellations, warranties, refunds
• The things that make you different, like "we only use X" or "we specialise in Y"

Write each answer the way you would say it to a customer, in plain language. You are not writing a legal document, you are briefing a new team member who will speak on your behalf.

Be specific, and be honest about the edges

Vague information produces vague answers. "We open in the morning" is useless. "We open at 7am weekdays and 8am Saturdays, and we are closed Sundays and public holidays" gives the assistant something it can actually say.

Just as important, tell the assistant what it should not do. If you never quote prices for big jobs over text because every job is different, say so, and the assistant will route those to a person instead of guessing a number. The clearer you are about the boundaries, the cleaner the handover to your team. This is the same line we draw in using AI to answer customer texts: the assistant handles what it knows and passes the rest to a human.

A worked example of the difference. A customer texts a physio clinic asking about a Saturday appointment. With a thin knowledge base:

Assistant: I'm not sure about Saturday availability, someone will get back to you.

With a proper one that includes hours, services and how booking works:

Assistant: Yes, we run Saturday mornings 8am to 12pm. Was this for an initial consult or a follow-up? Initial consults are 45 minutes and I can offer 8:30am or 10am this Saturday.

Same question, completely different customer experience, and the only thing that changed was what the assistant had been told.

Keep it current, or it turns on you

A knowledge base is not a set-and-forget document. The moment your prices change, your hours shift for summer, or you add a service, the assistant needs to know, or it will keep telling customers the old truth. An assistant confidently quoting last year's prices is worse than no assistant at all.

Build a habit around it. Whenever something changes in the business that a customer might ask about, update the knowledge base in the same breath. Tie it to things you already do: when you change your pricing, when you update your website, when you adjust your holiday hours, that is the trigger to update what the assistant knows. It takes a minute and it saves you from a hundred wrong answers.

A quick quarterly read-through helps too. Skim the whole thing, check nothing has drifted out of date, and add any new questions that have started coming up. If your team keeps flagging the same question that the assistant fumbles, that is a gap in the knowledge base telling you what to add next.

Feed it from the conversations themselves

Your best source of new entries is the incoming texts. The questions customers ask that the assistant could not answer are a ready-made to-do list. If you run a shared team SMS inbox, those handovers are all in one place, and reviewing them once a week shows you exactly where the knowledge base is thin.

Over time this creates a loop that makes the assistant steadily better. Real questions reveal gaps, you fill the gaps, and next time the assistant handles what it used to pass on. That is the whole reason two-way texting works so well with an assistant, and it is worth understanding the wider case for letting customers reply at all.

One realistic expectation to hold onto. A knowledge base makes the assistant accurate, not omniscient. When it genuinely does not know something, the right behaviour is to say so and hand over to a person, not to invent an answer. A well-fed knowledge base simply shrinks how often that happens, until the assistant is quietly handling most of what your customers ask. To see how the assistant fits the day-to-day across different trades, have a look at our use cases.