Ask a general AI model about your business and it will answer. It will sound sure. Sometimes it will be right. When it is wrong, it will be wrong in the same confident voice, and a customer will not be able to tell the difference.


For a small business, that is the real risk with AI. Not that it writes badly, but that it quotes a price you do not charge, a policy you do not have or opening hours from two years ago.


The fix has a name: grounding.


What grounding means


A grounded AI answers from facts you gave it, not from whatever it absorbed from the internet. Before it answers, it looks up the facts that apply. Then it writes the answer from those facts, and only those.


The usual technique is called retrieval-augmented generation, or RAG: find the right passages in your documents, hand them to the model and ask it to answer from them. It is a good start. On its own, it is not enough.


Where simple retrieval goes wrong


Retrieval finds text that looks like the question. That is not the same as the right answer.


  • Old versions. Last year's price list and this year's both match "how much is a tune-up." The model may pick the wrong one.
  • Near misses. A page about residential service matches a question about commercial service. The words fit; the answer does not.
  • Gaps. When the answer is not in your documents, a model will often fill the gap with something plausible instead of saying it does not know.
  • Facts in different places. Prices live in a spreadsheet, hours in your map listing, policies in a PDF. Searching one of them gives half an answer.

What a grounded setup adds


Four things turn retrieval into answers you can trust:


  • One current source of truth. Your facts in one place, with old versions retired, not five copies across five tools. That is the job of the Brain.
  • The right lookup for each fact. Prices from the price list, hours from the listing, policies from the policy page. Not one search across everything.
  • A source on every answer. Each answer can point to where it came from, so you can check it, and fix the source when it is wrong.
  • A clear "I don't know." When the facts are not there, the AI says so and passes the question to a person instead of guessing.

The goal is not an AI that knows everything. It is one that knows what it does not know, and acts on it.

Why it matters for AI search too


Grounding is not only about your own chat agent. ChatGPT, Perplexity, Gemini and Google's AI answers also try to base their answers on sources they can find. If your site states your facts clearly, in one place, with structured data, those engines have a better chance of citing you correctly. If your facts are scattered or out of date, they will cite someone else, or guess.


The same work helps with both: get your facts straight, keep them in one place and publish them plainly.


A test you can run this week


Take five questions customers ask you every week. Ask them to the AI tool you use, or the one you are thinking of buying. For each answer, check three things:


1. Is it right for your business today?

2. Can it show where the answer came from?

3. When you ask something it should not know, does it say so?


If it fails the third test, do not put it in front of customers yet.


How we handle it


At Humonix, agents answer from the client's Brain, and the chat agent on our own site answers from a published fact sheet. When the facts are missing, the agent asks a person instead of filling the gap. Anything public, priced or paid still waits for a named person's approval.