Skip to content
ChatLens360
Revenue & customer experience

The real cost of a chatbot hallucination

A hallucinated answer is not just an embarrassing glitch. It can cost a sale, a refund, a support hour and a customer. Here is how to size it up.

· Updated Nov 8, 2026

Illustration of a falling bar chart with coins slipping away for the article "The real cost of a chatbot hallucination", tagged Hallucinations, Revenue, Brand reputation.

When a chatbot invents an answer, people call it a hallucination. The word makes it sound harmless, almost funny. For a business, it is neither. A made-up answer delivered with confidence to a paying customer has a cost, and that cost is usually higher than it looks.

This post breaks the cost of a hallucination into parts you can actually estimate, and shows how to catch these answers before customers do.

What a chatbot hallucination is

A hallucination is an answer the chatbot produces that is not supported by your real information. It is not a typo or a formatting problem. It is the model filling a gap with something plausible.

Common examples:

  • A policy that does not exist ("We offer a 90-day price match guarantee").
  • A product feature you do not have.
  • A phone number, email address or URL that leads nowhere.
  • A delivery time, discount or fee that someone simply made up.
  • A confident answer to a question that should have been "I'm not sure, let me connect you with the team".

The danger is the confidence. A hallucinated answer reads exactly like a correct one.

One hallucination, many costs

Take this example from a home appliance retailer:

Customer: If I buy the washer today, is installation included? Chatbot: Yes! Free installation and removal of your old appliance are included with every washer purchase.

In reality, installation costs extra and old appliance removal is not offered at all. Here is what that one answer can cost:

  1. The awkward delivery. The customer expects installation. The delivery team arrives without it. Now someone has to decide whether to install it for free, charge for it, or leave.
  2. The support time. The customer calls, then emails, then calls again. Several staff members spend time on a problem the bot created.
  3. The concession. To keep the peace, the business often waives the fee or offers a discount. That is margin gone.
  4. The refund or return. If the customer feels misled, they may return the product entirely.
  5. The review. A customer who feels lied to is far more likely to say so publicly.
  6. The legal exposure. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline responsible for wrong information its chatbot gave about bereavement fares, and the airline had to compensate the customer. Your chatbot speaks for your business.

Estimating the cost: an illustrative model

Every business is different, so here is a simple model with clearly made-up numbers. Replace them with yours.

Say a single hallucination about included services leads to:

  • 45 minutes of combined staff time across support and operations, at $30 an hour: about $22.
  • A goodwill concession of $50.
  • A one-in-five chance the customer returns the product, losing a $100 margin: on average $20.

That is around $92 per incident before you count reviews or lost repeat business. If the same wrong answer is given to 30 customers a month, that is roughly $2,760 a month from a single wrong sentence.

The most expensive hallucinations are the repeatable ones. A bot does not make a mistake once. It makes the same mistake every time someone asks the same question in the same way.

Why hallucinations happen

Understanding the cause helps you fix it:

  • Missing information. The answer is not in your content, so the bot fills the gap.
  • Ambiguous content. Two pages say different things and the bot blends them.
  • Loose instructions. The bot was never told that "I don't know" is an acceptable answer.
  • Leading questions. Customers ask "Installation is free, right?" and an eager-to-please bot agrees.
  • Outdated sources. Old pages and PDFs describe policies that no longer apply.

A checklist for catching made-up answers

You can start reducing hallucinations today:

  • Ask about things you do not offer. "Do you do same-day delivery?" "Is there a student discount?" A good bot says no or says it does not know. A risky bot invents something.
  • Ask leading questions. "That includes free setup, right?" See whether the bot agrees just to please.
  • Check every specific detail. Prices, times, phone numbers, emails and links should all match a real source.
  • Look for conflicting content. Search your site for old policies and remove or update them.
  • Tell the bot what to do when unsure. Give it a clear fallback: admit uncertainty, link to the right page, or offer a human.
  • Re-test after every content or prompt change. A fix in one place can open a gap somewhere else.

Turning findings into fixes

For each hallucination you find, record three things: the question, what the bot said, and what the correct answer is with its source. That gives you a clear fix path. Either the content needs correcting, the content needs adding, or the bot's instructions need tightening.

If you want to see how this ties into the bigger picture, read how wrong chatbot answers quietly cost you customers.

See what your chatbot is making up

ChatLens360 generates realistic and tricky customer questions from your own website and documents, sends them to your chatbot, and flags every answer that is not supported by your facts. Each issue comes with the full conversation, the expected answer, the source page and a ready-to-paste fix prompt. Run a free chatbot audit or see how it works.

Put these ideas into practice

Run an automated audit of your own chatbot for free.