What a good AI assistant needs to know about your business

An AI assistant that knows nothing about your business guesses. And an assistant that guesses is worse than no assistant. What separates a useful solution from an embarrassing one is almost always what it has been told — not how advanced the model is.

Knowledge beats the model

It is tempting to think a better language model means better answers. In practice it matters less than people think. A top model with no knowledge of your business answers confidently and wrongly. A simpler model with the right information answers precisely.

The difference lies in the groundwork. It is the least glamorous part of an AI project, and the most important.

The actual answers

The most basic thing an assistant needs is the answers to what customers actually ask. Not a rough idea — the exact, current answers.

• Common questions and their definitive answers. Opening hours, delivery, returns, warranty, payment.
• Prices and terms. What applies now, not what applied last year.
• Product or service details. What you actually offer, and what you do not.

If these answers live only in the head of whoever has worked there longest, the assistant has nothing to learn from. The first step is then to write them down — independent of AI.

The voice and tone

An assistant represents the business. So it should sound like you, not like a generic robot.

It comes down to simple things. Are you formal or casual? Which words do you use for your products, and which do you avoid? A couple of examples of answers you are happy with teach the assistant more about tone than a long set of instructions.

Tone is not decoration. A customer quickly notices whether an answer came from someone who knows the business, or from a template.

The boundaries — what it should not answer

Just as important as what the assistant knows is what it should stay away from. A good assistant knows its own limits.

It should know which cases it should not try to solve: complaints with feelings involved, legal questions, anything to do with money or responsibility. Rather than guessing, it should pass such cases on to a human — with context.

An assistant that knows when it does not know builds trust. One that pretends to know everything loses it fast.

Where it gets its answers from

A good assistant does not guess from the internet. It pulls answers from your own sources: documents, routines, past cases. That is what lets it answer correctly about your business specifically, not about the industry in general.

This means the sources have to be tidy and up to date. If three versions of the price list sit in different folders, the assistant does not know which one applies — and then it answers wrongly with full confidence.

What happens when things change

A business does not stand still. Prices are adjusted, routines change, new products arrive. An assistant that was fed information at launch and never again becomes gradually outdated — quietly and imperceptibly.

So knowledge is not a one-off job. Someone has to own it: keeping the answers current, catching the cases nobody foresaw, and adjusting when something slips. It need not take much time, but it cannot be nobody's responsibility.

Start with what you already have

It sounds like a lot, but most of it already exists — scattered across emails, documents and people's heads. The groundwork is mostly about gathering and tidying it, not creating something new.

Often the most useful exercise is to write down the answers to the twenty most common questions. Do that, and you have done half the job before the technology is even mentioned. If you need help getting started, we are happy to have a chat.