How do I reduce AI hallucinations in customer-facing replies?

Practical ways for D2C brands to stop AI assistants inventing policies, prices or delivery dates, using grounding, limits, fallbacks and regular review.

Anadoris Digital · · 4 min read

Fog road
Photo: Frantzou Fleurine / Unsplash (CC0)

A hallucination is when an AI states something false as if it were true. In customer service, that might mean promising a delivery date the courier never gave, inventing a return policy, quoting the wrong price or claiming a product is suitable for a condition. The reply sounds confident and polite, which is exactly what makes it dangerous.

Language models generate likely-sounding text. When they lack the right information, they fill the gap with something plausible. Reducing hallucinations is mainly about making sure the right information is always present and that the AI is not asked to go beyond it.

Ground replies in approved sources

The most effective step is grounding: giving the AI the specific information it needs for each question, and instructing it to answer only from that. For a D2C brand, the sources are usually a short set of policy documents, product fact sheets, approved FAQ answers and a live, read-only connection to order and tracking data.

Keep these sources small, current and free of contradictions. If two documents give different return windows, the AI may pick either. Retire outdated versions completely, rather than leaving them in the folder.

Write sources in plain, unambiguous language. A policy that says returns are accepted “in most cases within a reasonable period” invites the AI to fill in details. A policy that states the window, the conditions and the exceptions clearly gives it nothing to invent. Clear writing for customers and clear writing for AI turn out to be the same thing.

Do not let the AI generate facts

Prices, order numbers, delivery dates, tracking links, discount codes and policy numbers should be inserted from your systems, not written by the model. A common design is to let the system look up the facts and pass them in, with the AI only phrasing the reply around them. If the lookup fails, the reply should say so and hand over, not guess.

Product questions deserve the same discipline. If a customer asks whether a cream suits sensitive skin and your fact sheet does not say, the correct reply is that you will check with the team, not a reassuring guess. Add explicit lines to fact sheets for the questions customers ask most, including what you cannot claim.

Set clear limits and fallbacks

  • Tell the AI exactly which topics it may answer and which it must hand over.
  • Instruct it to say it will connect the customer to the team when sources do not cover a question.
  • Forbid medical, legal or financial advice beyond your approved text.
  • Forbid promises about refunds, compensation or delivery dates not in the data.
  • Keep replies short; long replies give more room for invented detail.
  • Ask the AI to cite internally which source it used, so reviewers can check quickly.

Test before and after launch

Build a test set of real past questions, including tricky ones: questions about products you do not sell, policies you do not have, discounts that do not exist, and medical questions. Run the assistant against them before launch and after every change to prompts, tools or documents. A good assistant declines or hands over on the tricky ones rather than inventing an answer.

After launch, sample live conversations weekly. When you find a hallucination, trace it to the cause: a missing fact, a confusing document, a topic that should have been handed over, or an instruction the model ignored. Fix the cause, then add the question to your test set so it is checked every time.

Keep an eye on changes outside the AI as well. A new product, a policy update, a courier switch or a sale event can all create questions your sources do not yet cover. Update the sources before the change goes live, and add a few test questions about it, so the assistant is ready on day one.

A hallucination checklist

  • Small, current, consistent source documents with no outdated versions.
  • Facts such as prices, dates and order details inserted from data.
  • A clear list of topics the AI may answer and topics it must hand over.
  • A default fallback to a person when sources are silent.
  • A test set of tricky questions run after every change.
  • Weekly sampling of live replies, with each error traced to its cause.
  • An easy way for customers to reach a person at any point.

No setup removes hallucinations entirely. Plan for the occasional wrong reply with handover, monitoring and a quick way to correct customers when it happens.

Where to start

Collect fifty real customer questions, including ten awkward ones. Check that your source documents answer the normal ones clearly and that the awkward ones lead to handover. Run your assistant against the set and fix what fails before going live. Anadoris Digital builds customer assistants with grounding, data lookups and test sets in place from the start.

Related service: AI & Automation · More on this topic: AI & Automation articles

Common questions

Can a better AI model stop hallucinations?

Newer models are generally better at following instructions and admitting uncertainty, but none are free of hallucinations. Grounding, limits and review matter more than the choice of model for keeping customer replies accurate.

What should we do when the AI gives a customer wrong information?

Correct it quickly and honestly with the customer, honour reasonable expectations it created where appropriate, and fix the cause in your sources or rules. Add the case to your test set so the same error is caught in future.

Does asking the AI to be careful in the prompt help?

Clear instructions help, especially telling it to answer only from provided sources and to hand over when unsure. But instructions alone are not reliable; combine them with good sources, facts from data and regular testing.

Your next chapter

Let’s build what comes next.

A new website. A better system. A simpler way to work. Tell us where you want to begin.

Discuss your project