A vendor tells you the chatbot will be "trained on your business data." It sounds reassuring. It also raises fair questions. Is it reading everything you've ever written? Is your data now inside someone else's AI? And will it finally stop making things up?
To train a chatbot on your own data usually does not mean building a new AI from scratch. It means giving an existing language model a library of your approved material (pages, FAQs, price sheets, policies) and telling it to answer from that library. The bot knows what's in the library. It doesn't know anything else about your business.
That distinction explains most of what goes right and wrong with business chatbots, so it's worth a few minutes.
What does "trained on your data" actually mean?
There are two different things people call "training," and vendors often blur them.
Fine-tuning means adjusting an AI model's internal settings using lots of example conversations. It's expensive, slow to update, and rarely needed for a small business.
Retrieval is what most website assistants use. When a visitor asks a question, the system searches your material for the most relevant passages, hands them to the language model, and asks it to write an answer based only on those passages. A language model is the kind of AI that reads and writes everyday language. Retrieval is why you can change a price on Monday and have the bot quote the new one on Tuesday, with no retraining.
When a vendor says "trained on your data," ask which of the two they mean. For most small businesses, retrieval is the right answer, and it's the easier one to keep accurate.
What can you feed it?
The material you supply is usually called a knowledge base. Typical sources:
- Your website pages (services, about, pricing, locations)
- A written FAQ
- PDFs such as price lists, intake forms, or policy documents
- Hours, service areas, and holiday closures
- Short notes on how you want it to behave ("always offer to book a visit," "never quote a final price")
Quality matters more than quantity. A short, current, clearly written FAQ beats a folder of old brochures that contradict each other. If two documents disagree, the bot may pick either one.
What will the chatbot never know?
This is the part that surprises owners. A bot trained on your material does not know:
- Anything you haven't written down. Your unwritten rule that you don't take jobs past a certain distance is invisible to it.
- What happened today. It doesn't know you're out sick or that a supplier delayed an order unless someone updates it.
- Private details about individual customers, unless it's connected to your CRM (customer records system) on purpose and with permission.
- Whether a customer is telling the truth. It can't verify a warranty claim or a past purchase on its own.
- Your judgment. Exceptions, discounts, and tricky situations belong to you.
A well-built assistant is also told what to do when it hits one of these gaps: say it isn't sure and offer to pass the question to a person. If a vendor can't show you that behavior, treat it as a warning sign. We cover other limits in What an AI Chatbot Can and Can't Do for a Small Business.
Does your data end up inside someone else's AI?
It depends on the vendor and the plan, which is why you should ask. Questions worth putting in writing:
- Is my material used to train or improve models that other customers use?
- Where is it stored, and who at your company can see it?
- Can I delete it, and how fast?
- What happens to conversation logs, and how long are they kept?
- Does the bot show visitors where an answer came from?
Don't upload anything to a chatbot knowledge base that you wouldn't put on your public website, unless the vendor has answered those questions to your satisfaction. Internal pricing formulas and customer lists are the usual examples.
A simple readiness exercise
You can test your own material before paying anyone. Set aside an hour.
- Pull the last 20 questions customers asked by email, phone, or form.
- For each one, find the exact sentence in your current website or documents that answers it.
- Mark each as found, unclear, or missing.
Say you end up with 12 found, 3 unclear, and 5 missing. That tells you the bot could handle about 60 percent of those questions today, and you know precisely which eight to fix first. This is an illustration, not a benchmark. Your numbers will differ, but the exercise always shows where the gaps are, and fixing them helps your website too.
A chatbot is only as good as the material behind it, so write the answers first.
Keeping the knowledge base accurate
Treat it like a shared document, not a one-time setup.
- Give one person ownership of updates.
- Review it when prices, hours, or policies change.
- Read a sample of real conversations each month and note wrong or vague answers.
- Add the questions it couldn't answer to your list of things to write.
This is also where a chatbot quietly pays for itself: the unanswered questions are a free list of what your customers can't find on your site. For more on what happens to those conversations, see How an AI Assistant Qualifies Leads Before They Reach Your Inbox.
How Deca Digital helps
We build AI assistants for websites that answer from your own approved content, ask for details when needed, and hand off to a person when they aren't sure. We help you gather and tidy the source material, set the handoff rules, and plan how it stays current. We're based in Celina, Texas, work with businesses across North Texas, and also work remotely. You can see what we offer on our services page.
If you'd like to know what your chatbot would be able to answer today, book a free discovery call.