Explainer

How to let an AI assistant answer visitors without it making things up

The real risk with an AI chat on your site isn't that it's rude — it's that it confidently invents answers. Here's how grounding, limits and handoff keep it trustworthy.

Key takeaways
  • A general AI model will happily guess — on your website, a wrong guess about price or availability is worse than no answer.
  • The fix is grounding: the assistant answers only from your own content, not from what the model 'knows'.
  • Clear limits and a clean handoff to a human are features, not failures — they're what makes the whole thing trustworthy.

The thing that stops most businesses from putting an AI assistant on their website isn’t tone or technology. It’s a reasonable fear: what if it tells a customer something that isn’t true? A confident, wrong answer about your prices, your availability or your policies isn’t a small glitch — it’s a promise you didn’t make, given to a customer in your name.

That fear is healthy. And the good news is that the fix is well understood. A trustworthy website assistant isn’t one that knows everything; it’s one that’s built not to make things up.

Why a general model guesses

Large language models are trained to be helpful and fluent, which means their default behaviour is to produce a plausible answer to almost anything. Ask a general model “how much does your deluxe package cost?” and it won’t say “I don’t know” — it’ll generate something that sounds right, because sounding right is what it was trained to do.

On a public website, that’s exactly the wrong instinct. You don’t want fluent guesses about your business. You want the truth, or an honest “let me get someone who can help.”

Grounding: answer only from your content

The single most important safeguard is grounding. Instead of letting the model answer from its general training, you restrict it to your material — your pages, your services, your prices, your FAQs, your policies. When a visitor asks something, the assistant retrieves the relevant piece of your content and answers from that, not from what the model happens to have absorbed.

The practical effect: it can tell a customer what’s actually true about your business, and when the answer isn’t in your content, it knows it doesn’t know. That’s the difference between an assistant that represents you and one that improvises as you.

The test to apply: ask any assistant you're considering something your website doesn't cover. A trustworthy one admits it doesn't know and offers to connect you. A risky one invents a confident answer. Try it before you trust it with customers.

Limits and handoff are features

It’s tempting to judge an assistant by how much it can answer. For a business, the better measure is how gracefully it handles what it can’t. A good assistant has visible edges:

  • It declines to guess. When something isn’t in your content, it says so rather than filling the gap.
  • It hands off cleanly. For a complicated, sensitive or high-value question, it brings in a human — ideally capturing the context so your team isn’t starting cold.
  • It stays in scope. It’s there to help with your business, not to be talked into off-topic answers it has no business giving.

Far from being weaknesses, these limits are what make the whole thing safe to put your name on. An assistant that knows when to stop is one you can actually trust in front of customers.

Trust is the point, not a trade-off

There’s a myth that you have to choose between an assistant that’s helpful and one that’s safe. You don’t. Grounding makes it more useful, not less — because answers people can rely on are the only answers worth giving. A visitor who gets a correct, specific answer and a clean handoff when needed trusts your business more, not less.

That’s the principle Voxy is built on: it answers from your own content, so it says what’s true about your business; it books appointments and captures leads within that; and it brings in your team the moment a person is genuinely needed — in nine languages, without ever pretending to know what it doesn’t. You get the speed of instant answers and the safety of never putting words in your mouth. It’s free to try for 14 days, no card required.