Why AI chatbots make things up, and how citations fix it

An AI agent that invents your policies is worse than having no agent at all. In 2024, an airline's chatbot invented a bereavement refund policy, and a court ordered the airline to honor it. The bot did what language models do by default: it produced the most plausible-sounding answer, and the plausible answer was wrong.

If you're putting AI in front of your customers, plan for that exact risk: an agent that answers confidently and wrongly, in your brand's voice, about your own policies, and sounds helpful while it does it.

Why models invent answers

Language models are trained to continue text convincingly. Ask one something outside what it knows and it fills the gap with the statistically likely shape of an answer. That's fine when you're brainstorming. It's a disaster when the question is about your returns policy.

Vendors often answer this with 'better prompts' or 'guardrails', which are instructions layered on top of the same mechanism. They make invention less frequent. They can't remove it, because the model's underlying objective, to always produce an answer, hasn't changed.

Retrieval with receipts

The fix is in how the system is built, and it has three parts. Restrict the agent to answering from a specific body of content, retrieved at question time: your website, your docs, your policies. Require it to cite the passage behind each claim. And when retrieval comes back empty, make the correct output 'I don't know'.

Citations do two jobs at once. Your visitor gets an answer they can check by clicking through to the source. You get an auditable record of every conversation: when an answer looks off, the citation shows exactly which page produced it, and usually the page is what needs fixing.

"I don't know" is a feature

Visitors trust an assistant more after they have seen it decline once, because the refusal proves the confident answers mean something.

Every refusal is also logged, which gives you a precise list of gaps in your content. Write the missing page, the next crawl picks it up, and the same question gets a cited answer the following week. The agent answers or logs the gap, and each round makes the website and the agent better.

How Agentmatica answers a policy question

Agentmatica is built on this design. A shopper asks, 'Can I return something I bought in the sale?' The agent pulls the answer from your returns page, replies in plain words, and links the page it used so the shopper can read the policy for themselves. If your returns page doesn't mention sale items, the agent says it doesn't know and offers to bring in a person. Your team takes over the live chat, answers, and hands the conversation back to the agent when they're done.

You stay in charge of what it says. Private directives let you tell the agent how to handle topics you care about, and the gaps it logs show you which pages to write next. The same grounding carries into sales: when a shopper asks which product fits, the agent answers from your catalog and shows the matching products as cards.