How to Train an AI Agent on Your Business (What to Feed It)

The most common thing that goes wrong when a business puts an AI agent on its phone or website is not the technology. It is that somebody pointed the agent at the company website, pressed go, and assumed the machine would work the rest out.

It will not. An AI agent answering for your business is only as good as the material you give it, and your website is written to sell, not to answer. When you train AI on your business properly you are writing something closer to an induction pack for a very fast new hire who will never ask you a clarifying question, never guess correctly about something you left out, and never quietly check with a colleague before speaking to a customer.

The good news is that the pack is smaller than you think. For most service businesses, six ingredients cover the overwhelming majority of real conversations.

Start from the twenty questions you already answer every week#

Before you write anything, spend twenty minutes listing the questions you and your team answer over and over. Not the ones you wish people asked. The actual ones.

The fastest way to build this list is to raid your own records. Scroll your last hundred text messages with customers. Open your email and read the first line of the last fifty enquiries. Read your Instagram and Facebook DMs. Ask whoever answers the phone to keep a tally sheet for three days.

You will find the same shapes repeating. Do you come out to my area. How much is it roughly. How soon can you get here. Do you charge for a quote. Do you do Saturdays. What do I need to do before you arrive. Are you registered. Do you take card.

Write the real answer to each one, the way you would say it, in two or three sentences. That document is the single highest-value asset in the whole build. Twenty properly written answers will resolve more conversations than four hundred pages of website content, because they map onto what customers actually type.

Six things you need to train AI on your business#

Once the question list exists, fill in the surrounding structure. This is the whole of a workable AI knowledge base setup for a typical service business.

1. A current service list with prices or price bands. Exact prices where you have them, honest ranges where you do not, and a clear note on what triggers the top of the range. If you refuse to give the agent any pricing, it will deflect every pricing question and customers will leave.
2. Your service area, written as suburbs or a radius rather than a vague region name. "Auckland" is not usable. A list of suburbs plus a travel surcharge rule is.
3. Hours, including public holidays, and what happens outside them. If the agent runs overnight, it needs to know what an after-hours customer should expect, which ties into how you handle after-hours enquiries.
4. The twenty answers you just wrote.
5. Escalation rules. Which topics the agent must hand to a person, and what it should say while doing it.
6. A short tone note. Two or three lines about how you sound, plus a handful of words you never use.

That is it. Every hour you spend beyond this on more source material has a lower return than an hour spent reading transcripts after launch.

Why dumping your whole website makes answers worse#

Vendors love the site-crawl demo because it looks like magic. Paste a URL, watch the agent answer questions about the business thirty seconds later. It is a good demo and a bad configuration.

Marketing copy is optimised for persuasion. "We pride ourselves on prompt, professional service" contains no retrievable fact. When a customer asks how quickly you can get to Papakura, the agent retrieves the sentence with the highest surface similarity, and a page full of adjectives gives it nothing to work with. It then does the thing you least want, which is to produce a confident sentence assembled from atmosphere rather than information.

Old content is the second problem. Most business sites carry a pricing page from two years ago, a services page that lists something you stopped offering, and a blog post with a phone number you no longer use. A human reader ignores the stale bits by instinct. A retrieval system treats them all as equally true, and the contradictions surface at the worst moment.

The third problem is dilution. Every irrelevant page you add increases the chance that the retrieval step pulls the wrong chunk. Adding your About Us page and your careers page to the knowledge base does not make the agent smarter about drainage.

If you already have website content that answers questions well, use it. Copy the passages that contain real facts into your knowledge base as clean text. Leave the rest out.

Write escalation rules before you write anything clever#

The rules that stop an agent doing something are more valuable than the rules that let it do something. Write them early, in plain language, and be specific about the trigger rather than the category.

For a typical trades or clinic setup, the list usually includes anything involving a complaint about work already done, anything where a customer says they want to speak to a person, any question about liability or insurance, any clinical or legal question, and any quote above a value you nominate. The mechanics of doing this without infuriating the customer are covered in the handoff, and it is worth reading before you write your rules rather than after.

Alongside escalation, write the refusals. Tell the agent explicitly what to do when it does not know. The correct behaviour is to say it does not have that information and offer to take a message or connect a person. The failure mode is an agent that invents a plausible answer because nothing told it that "I don't know" was an acceptable output.

Feeding it your calendar and your systems#

Knowledge is only half of it. An agent that can answer but cannot act is a well-read voicemail. To book, it needs a live connection to your calendar and a set of rules about how your day works.

Those rules are more specific than most people expect. How long each job type takes. How much travel time to leave between jobs in different suburbs. Which slots you keep free for emergencies. Whether a new customer can book the last slot on a Friday. What happens when someone tries to book a two-hour job into a thirty-minute gap. Getting these wrong produces the exact failures described in letting AI book appointments, and they are embarrassing in front of a customer in a way that a wrong price is not.

If the agent also handles social channels, the same knowledge base should drive them, so a question asked in an Instagram DM gets the same answer as the same question asked on the phone. Keeping one source of truth across channels is most of the work in automating WhatsApp and Instagram DMs.

The fortnight that actually creates accuracy#

Configuration gets you to roughly right. Transcripts get you to reliable. You cannot train AI on your business in a single sitting and never revisit it, and the owners who expect to are the ones who end up disappointed.

Run the agent in shadow mode first if your setup allows it, meaning it drafts answers that a human reviews and sends. A few days of this exposes the gaps cheaply. Where shadow mode is not practical, launch on one channel only, usually web chat, where a wrong answer is recoverable.

Then read every conversation for the first two weeks. You are looking for three things, and they show up quickly. Questions the agent could not answer, which become new entries in the knowledge base. Answers that were technically correct but not how you would have said it, which become tone corrections. And moments where the agent should have escalated and did not, which become new rules.

Most of the accuracy improvement in an AI agent comes from this loop rather than from anything in the initial build. A business that spends a week perfecting its source documents and then never reads a transcript ends up with a worse agent than one that launched with a rough knowledge base and corrected it twenty times in a fortnight.

After the first two weeks, drop to a weekly skim, then monthly. Add a standing rule that whenever a price changes or a service is added, the knowledge base gets updated in the same breath as the website. A stale knowledge base decays faster than you would think.

What to do next#

Everything else in the effort to train AI on your business follows from one document, so start there. Open a blank page and write the twenty questions. Do not start with software, do not start with vendor calls, and do not start by trying to work out how to train a chatbot in a technical sense. The list of real questions is the thing that determines whether any of this works, and it is the one part nobody can do for you.

Once you have it, you will also have a fairly clear view of how much of your enquiry volume an agent could genuinely resolve, which is the number that should drive the rest of the decision. For how the pieces fit together end to end, including pricing and setup time, the AI receptionist for small business guide covers the full picture.