Try it free
Product
Everything you can hand overSee every feature →
Set it up by talkingJust talkTell it what happened. It does the work.Set up by talkingIt arranges itself from a conversation.Your kind of workIt takes the shape of your work.Your wordsYour words, not ours.Your processPipelines and stages you design.Your questionsThe same questions, asked the same way.
Messages and answersInboxEvery message that reaches you, answered.Hand-offWhat it can't answer comes to you, with the reason.KnowledgeIt knows how you work.Web researchIt gathers from the web.
Promises and follow-upAfter the callSay it once, after the call.PromisesEvery promise, on its date.Going quietIt notices what's going quiet.AutomationsRoutines that run themselves.
Records and numbersRecordsEveryone's whole story, in one look.ImportBring what you already have.Products & quotesWhat you sell, and what you quoted.DashboardsYour numbers, from day one.AskAsk your work anything.
ControlTeamYour team, each with their part.ApprovalsJudgement waits for you.Reasons & sourcesNothing made up. Everything with its reason.
What’s new →
Use cases
What Felesh looks like for youAll use cases →
For you aloneYour own daysA business ownerA creatorSomeone who decides
TeamsA recruiterA project managerA product managerA support deskA B2B sales team
BusinessesA local grocery shopA training instituteAn online shopA small businessA wholesalerA recruitment agencyAn accounting firmA law officeA local clinicA dental clinic
Large organisationsAn enterpriseA hospitalA holding group office
PricingTrust
Resources
BlogGuidesChangelogAcademyAbout
Log in

How do I keep a customer chat from giving wrong answers?

A chat gives wrong answers when it fills gaps. Teach it what to say, let it admit what it does not know, and hand the rest to a person.

3 min read · · The Felesh team

Why a chatbot gives wrong answers

A customer asks about a discount you never offered. The chat has nothing on it, so it makes one up. It sounds reasonable, the customer acts on it, and you clean up the mess. Chatbot wrong answers usually have one cause: the tool is built to say something, even when it has nothing to say. A chat that cannot say “I do not know” will guess.

So the fix is not a cleverer chat. It is a chat that answers only from what you have taught it, and a route to a person for everything else.

Fig. 1 A chat answers from what you taught, never guesses, and hands the rest to you with the reason.

How to keep a chatbot from giving wrong answers

Answer only from what you have taught it. Every reply should come from the documents, pages and written instructions you provided. If the question matches nothing in that material, the chat should say it does not know rather than attempt an answer.

Keep the material current. If the chat still holds last year’s return window, it will give a wrong answer. Review what you have taught it when prices, hours or policies change.

Test with real questions. Do not test with what you think customers will ask. Pull up your actual messages, feed those phrases to the chat, and look at what comes back. A wrong answer is a gap in the material: fill it. A confident wrong answer where there was no material is a chat that guesses, and that needs fixing at the source.

When the chat should hand over to a person

The chat hands over when the question falls outside what you taught it. It also hands over when the customer asks for a person.

A good hand-over carries the reason, so you know what the customer asked and why the chat could not answer. See hand-off to a person, with the reason.

How this goes on Felesh

  1. Teach it. Upload your documents and write the instructions you want customers to be told. Everything the chat may say comes from there.
  2. It answers from that material. A customer’s question is answered from what you taught it. If nothing fits, it says it does not know instead of guessing.
  3. It hands over with the reason. The conversation goes to Needs you with the reason it could not answer. Nobody is notified, so it is a list you look at.
  4. You read it and reply. You open the thread, see what was asked, and answer.

Felesh knows what you put on it: the documents and instructions you give it, the contacts and companies you import for review, and the work done on it. It does not read your mail or your drives, and it does not connect to other tools.

What you still check

The material itself. If you upload a page with an old price, the chat repeats the old price. Felesh cannot know what is true for your business; you set the facts.

The Needs you list. Every hand-over shows a question your material could not answer. If several customers ask about the same thing, add it to what you taught it. If they ask for something you do not offer, say so in the material. The list is the best guide to what is missing.

Keep it up to date. Review the material when your business changes, and before each busy season.

A chat your customers can rely on answers only what you taught it and hands over when it cannot. See how Felesh avoids made-up answers.

Next question Do I have to tell customers they are talking to an AI? What Article 50 of the EU AI Act says about telling people they are dealing with an AI, and who the article addresses.