How to train an AI on your business's services, pricing, and FAQs

When someone says "we need to train the AI on our business," what they usually mean is: get the AI to a point where it can answer the questions customers actually ask, with information that is accurate for this specific business. That is a documentation task with a technology layer on top of it, and the documentation is the harder part for most business owners.

The overview of what AI can actually do for a service business right now covers why the knowledge base is the foundation everything else depends on. This article covers the practical steps to build one that works.

Step 1: gather the information your AI actually needs

Before anything else, the business needs to pull together the information the AI will draw from. This is the stage most people underestimate, partly because the information feels obvious when you know the business well. Writing it down in a form someone else could use, which is essentially what the AI requires, reveals how much is assumed rather than explicit.

The information falls into four categories.

Services. Every service the business offers, described in enough detail that someone with no prior knowledge of the business could understand what it includes and what it does not. This means naming the service, explaining what is involved, specifying what is included in the standard offering, and noting what is excluded or treated as an add-on. Vague service descriptions produce vague AI answers.

Pricing. This is the category businesses most often avoid documenting. The instinct to keep pricing out of the AI's knowledge base on the grounds that "it depends" is understandable, but it produces an AI that deflects every pricing question, which frustrates customers. The solution is not to give a single fixed price where one does not exist, but to give an honest range, note the main factors that affect where in the range a job lands, and give a clear indication of how the customer can get a more precise figure. A range is more useful than silence.

Operations. How the business works: coverage areas, typical lead times, how bookings are made, what happens after a quote is accepted, cancellation terms, and what customers should expect after they contact the business. These are the questions customers ask because they are trying to plan around the service, and AI that cannot answer them creates friction rather than removing it.

Common questions. Pull the last two months of enquiries and look for the questions that appear most often. Not the questions the business thinks it gets, but the ones that actually show up in messages, emails, and phone records. These real questions are the most useful source of content for the FAQ section of a knowledge base.

Step 2: structure your knowledge base so the AI can use it

A knowledge base written for a human reader and a knowledge base optimised for AI use are different documents. A human reader follows narrative; an AI retrieves information from structured entries. The structure that works best for most AI platforms on the market is a series of question-and-answer pairs, organised by topic.

Each entry should be self-contained. Rather than writing "see above for pricing" or "as mentioned in the services section," each entry should include the full relevant information, even if that means some repetition. When an AI retrieves an entry to answer a customer question, it is not reading the full document. It is pulling the most relevant entry, and that entry needs to stand on its own.

Tone matters in a knowledge base. The AI will reflect the register of the knowledge base content when constructing responses. Formal, corporate-sounding entries produce formal responses. Natural, conversational entries produce responses that feel closer to how the business actually communicates. Writing the knowledge base in the voice the business uses with customers is not an optional refinement; it directly affects how the AI comes across.

EveryCatch's follow-up sequences depend on a well-structured knowledge base for the same reason: the follow-up messaging pulls from the same information the AI uses for first response, and inconsistency between what the AI said initially and what the follow-up says creates confusion. A single, well-maintained knowledge base feeds all of these outputs.

Step 3: test the AI against real customer questions

Before the AI goes anywhere near a live customer, it needs to be tested against the actual questions customers ask. Not hypothetical questions, but real ones pulled from the business's message history and phone notes.

The test process is straightforward. Take twenty to thirty questions the business has received in the last few months and ask the AI each one. For each response, assess whether the answer is accurate, whether it matches the tone the business wants to project, and whether it would leave the customer with what they needed or with more questions. Any response that fails one of those three checks indicates a gap in the knowledge base.

Common gaps that testing surfaces: pricing questions where the AI gives an unhelpful non-answer; service-scope questions where the AI either under or overpromises what is included; availability questions where the AI gives a general answer when a specific process exists; and handoff questions where the AI does not know what to do when it cannot fully resolve something. Each gap is a knowledge base entry that needs adding or rewriting.

Testing also catches tone problems that are not obvious when writing the knowledge base. An AI that sounds noticeably different from how the business normally communicates is a problem even when its answers are accurate. If the test responses feel wrong, revising the tone guidance in the knowledge base usually fixes it.

Step 4: keep the AI knowledge base current as the business changes

A knowledge base that is accurate on day one and never updated becomes a liability. Prices change. Services are added or discontinued. Coverage areas expand. Booking processes are updated. An AI operating from a knowledge base that does not reflect current reality will confidently give customers incorrect information, which is harder to recover from than the AI simply not having an answer.

The simplest way to maintain a knowledge base is to assign ownership. Someone in the business should be responsible for checking the knowledge base whenever a material change happens, updating the relevant entries, and reviewing it on a quarterly basis even when nothing obvious has changed. Quarterly reviews often catch drift that would not prompt an immediate update: a price range that was set when costs were different, a service description that no longer reflects how jobs are done, an FAQ answer that was provisional and was never revisited.

From the EveryCatch team

EveryCatch works through the knowledge base build as part of the onboarding process, testing it against real questions from the business's own enquiry history before going live.