An AI assistant that answers your customers gives wrong answers when it has to guess. It guesses when it lacks the facts, when its facts are out of date, or when it is asked something outside its job. Each cause has a fix: give it one set of approved facts, tell it what it may not answer, let it say it does not know, hand over to a person at set points, and test it before and after launch.
Why do AI assistants give wrong answers?
Language models write the most likely answer to a question, and that answer is not always true. Anthropic, which makes Claude, puts it plainly: even the most advanced models can sometimes generate text that is factually incorrect. In a customer-facing assistant, wrong answers usually come from five places:
- Missing facts. Asked about a service it has no information on, it fills the gap.
- Old facts. Last year’s prices or opening hours are still in its instructions.
- Questions outside its job. A booking assistant asked for legal or medical advice.
- Pressure to help. Instructions such as “always answer the customer’s question” push it to answer when it should not.
- Vague instructions. “Be helpful and friendly” says nothing about what it may promise.
Give it one source of truth
Write down the facts the assistant may use, in one place: services and what each includes, prices or how they are worked out, opening hours, locations and areas served, booking rules, cancellation and refund policies, and answers to the twenty questions customers ask most.
Give each section an owner and a review date. When a price or a policy changes, the owner updates this one document and the assistant uses the new version. Do not let it answer from general knowledge about your industry. Instruct it to use only the approved facts, and to quote them where it can. Anthropic recommends grounding answers in direct quotes from the source material and checking claims against it as ways to reduce wrong answers.
Tell it what it may not answer
List the topics it must not handle, and what it should do instead. Typical examples:
- Advice that needs a qualified person: legal, medical, financial or safety advice.
- Prices for custom work, beyond the ranges you publish.
- Exceptions to your policies, such as refunds outside the stated terms.
- Complaints and disputes.
- Anything about another customer.
A short, explicit list works better than a general instruction to “be careful”.
Let it say “I don’t know”
The most effective single change is permission to admit uncertainty. Anthropic’s guidance on reducing hallucinations starts with allowing the model to say it does not know. Give the assistant the exact words to use, such as “I don’t have that information, so I’ll pass this to our team”, and what happens next. A customer who hears “I’ll check” trusts you more than one who is told something wrong.
Decide when a person takes over
Write down the moments when the conversation goes to staff:
- The customer complains, is unhappy, or asks for a person.
- Payment, a refund or a cancellation fee is involved.
- The matter is urgent, such as a leak, a medical issue or a same-day booking problem.
- The same question has been asked twice without an answer.
- The topic is on the “may not answer” list.
Pass the whole conversation to the person who takes over, so the customer does not repeat themselves, and tell the customer when they will hear back.
Test it before customers do
Collect fifty to a hundred real questions from your inbox, chat history and phone notes. Include the awkward ones: questions with a wrong assumption (“you’re open Sundays, right?”), questions about services you do not offer, and attempts to get a discount. Run them all, mark each answer right, wrong or should-have-handed-over, and fix the approved facts or the instructions behind every wrong one. Run the full set again after every change.
Keep watching after launch
Read a sample of real conversations every week for the first two months, then every month. Look for three things: answers that were wrong, questions it could not answer, and hand-overs that should not have happened. Questions it could not answer are the best list of what to add to the approved facts.
Protect customers’ data
Collect only the details you need to help or book. Tell customers they are talking to an AI assistant, and how to reach a person. Follow the data protection law where you operate, and check where your AI provider stores conversations and whether it uses them for training. Most business plans let you switch that off.
Questions to ask a provider
If someone else is setting up your assistant, ask:
- Where do its answers come from, and who can update them?
- What does it do when it does not know?
- When does it hand over to a person, and how?
- How was it tested before launch, and how often is it reviewed?
- Where are conversations stored, and are they used for training?
A provider who cannot answer these clearly is asking you to trust the assistant without checks. Our guide to AI automations worth starting with covers where an assistant fits among other automations, and choosing an AI model compares the models behind them. Our AI Front Desk answers only from the facts you approve, in English, Mandarin or Malay, and hands anything unsure to a person.