Stop Your Support Chatbot From Hallucinating: 7 Practical Fixes

A support chatbot starts giving wrong answers the moment it’s allowed to guess. Ground it in approved sources, pull live account data from your systems, and treat “I don’t know—let me get a person” as a valid outcome.

This post breaks down why models invent details and walks through a 7-step fix, plus accuracy testing and two public failures.

Your support chatbot gives wrong answers because nothing stops it from guessing. Ground it in your own documents and connect it to live systems for account data. Then make handing off to a person an acceptable outcome.

Even when the source is right in front of them, models still invent details. On Vectara’s hallucination leaderboard (May 2026), models asked to summarize a document using only that document added unsupported facts 1.8% of the time at best.

Large language models write the most likely next words. They have no built-in check for whether a sentence is true. OpenAI’s September 2025 research on hallucinations adds a second reason: accuracy-only tests give no credit for “I don’t know”, so models learn that guessing pays.

The same paper shows the cost of that habit. On one fact test, a model that declined 52% of questions was wrong on 26% of them. A similar model that almost never declined scored slightly higher on accuracy, but was wrong 75% of the time.

In a support setting, the model is rarely the only cause. Wrong answers usually enter at one of these points:

Anatomy of a reply: where wrong answers get in

After a death in the family, Jake Moffatt asked Air Canada’s website chatbot about bereavement fares. It said the reduced fare could be claimed within 90 days of the ticket being issued. The airline’s real policy did not allow claims after travel.

In February 2024, British Columbia’s Civil Resolution Tribunal ruled against Air Canada. It rejected the airline’s argument that the chatbot was responsible for its own actions. The tribunal found that “Air Canada did not take reasonable care to ensure its chatbot was accurate,” and ordered it to pay C$812.02. The sum was small. The lesson was not: your company answers for what its chatbot says.

In April 2025, users of the coding tool Cursor kept getting logged out when they switched computers. An AI support agent answering email told them this was expected behavior under a one-device policy. There was no such policy. The logouts were a bug. Co-founder Michael Truell apologized publicly, the user was refunded, and AI email replies are now labeled.

Both bots answered a policy question with confidence and no source. Both incidents would have been a non-event if the bot had said it wasn’t sure and passed the case to a person.

Customers already expect this risk. In a Gartner survey of 5,728 customers published in July 2024, 64% said they would prefer companies didn’t use AI for customer service. AI giving wrong answers was one of their top three concerns.

For e-commerce merchants, teams at https://neuradynamics.ai/ have built AI assistants that draft replies from Shopify order data. An agent approves each one before it is sent. For a fintech company, an omnichannel support system answers routine email and routes complex cases to the team. Both answer from approved sources and put a person in the loop.

This is the order we would work in for any AI chatbot for customer support. The early steps are cheap and fix most problems. The later ones matter most for refunds and billing.

Steps 3 to 6 in one picture: every question takes one of three routes

Tip: Write the decline message yourself. “I don’t want to give you the wrong refund information, so I’m passing this to a colleague” keeps trust. A generic error message loses it.

Once you have tagged your bad answers, use this table to pick the fix. Effort is our rough estimate for a typical help desk setup.

Build a test set from real tickets, not from questions your team invents. Our recommendation is about 200 questions, each with the correct answer and the document it comes from. Include questions your documents don’t cover, outdated policies, angry customers and attempts to talk the bot into a discount.

Score each reply as correct, correct but unsupported, rightly declined, or wrong. Track the wrong rate on its own. OpenAI’s numbers show why: a single accuracy score hides a model that is wrong nearly three times as often.

Declining vs guessing: similar accuracy, very different error rates

Decide the pass mark before you test. For refund, billing and legal questions, accept zero wrong answers in the test set, even if that means more declines. Rerun the full set when you switch models, because error rates vary widely even within one vendor’s line-up.

The Air Canada ruling says yes. The tribunal treated the chatbot as part of the airline’s website, so the airline owned its statements. Hold your own AI customer support to the same standard.

In India, the Reserve Bank of India published its FREE-AI framework in August 2025, with 7 principles and 26 recommendations for AI in financial services. If you serve customers of a bank, NBFC or fintech, build in disclosure, a human route and a log of bot replies.

Note: This is not legal advice. Check the rules for your sector and country with your legal team.

Language models predict likely words and have no built-in fact check. OpenAI’s 2025 research adds that training and testing reward guessing over admitting uncertainty. In support, the trigger is usually a gap—missing or outdated document, a search that pulls the wrong passage, or a question needing live account data.

No. You can make them rare and catch them before customers see them. A 2024 Stanford study found legal AI tools grounded in their own databases still hallucinated on 17% to 33% of queries. The practical goal is a bot that declines or hands off when unsure, plus checks on high-stakes answers.

It reduces them a lot, but not to zero. Retrieval-augmented generation makes the bot answer from your documents instead of memory. It still fails when documents are wrong, when search pulls the wrong passage, or when the model adds details. Pair it with citations, declines and testing.

Use a fixed set of real customer questions, each with a known correct answer and source. Score replies as correct, unsupported, rightly declined or wrong, and track the wrong rate separately. Rerun the set after every change to content, prompts or model, and review flagged live conversations every week.

In the Air Canada case, yes. In February 2024, a Canadian tribunal rejected the airline’s claim that its chatbot was a separate entity. It ordered Air Canada to pay C$812.02 for a fare rule the bot invented. Treat every chatbot answer as a statement from your company, and check local rules with a lawyer.

Yes, when its sources don’t cover the question. A clear “I’m not sure, let me get a colleague” costs one handoff. A confident wrong answer can cost a refund, a complaint or a legal claim. In Gartner’s 2026 survey, 87% of customers said companies using GenAI for customer service must provide access to a human agent.

A customer service chatbot that gives wrong answers usually has a setup problem you can measure and fix. Start this week: pull 30 days of conversations and tag each bad answer by cause. That list tells you whether to fix your documents, your search or your handoff rules first.

Originally published at https://neuradynamics.ai/blogs/customer-service-chatbot-wrong-answers