When a chatbot answers a sales enquiry, a wrong price or an appointment that does not exist can leave a potential customer with the wrong expectations. Start by tracing one mistake back to the information and rules behind it. Make sure your team can take over with the customer's question and relevant details.
This guide is for a chatbot your business runs on its website or uses to prepare customer replies. The seven checks below help you and your supplier diagnose the problem. Examples are fictional and illustrate what to test.
Why can a chatbot sound certain when it is wrong?
Generative AI can produce convincing text without getting the facts right. A confident tone is not evidence that the system checked your prices, policies or calendar.
The US National Institute of Standards and Technology describes confidently presented false content as “confabulation”, also called hallucination, in its Generative AI Profile. It identifies this as a risk of how generative systems produce outputs.
A wrong reply can also repeat an outdated document or use the wrong service description. Those problems need different fixes. Asking the chatbot to “be more accurate” does not tell you which one occurred.
1. Can you identify exactly what went wrong?
Record the customer's question, relevant conversation context and the incorrect reply. Then write down the correct answer and the source that supports it.
Mark the specific error: wrong fact, missing condition, unsupported promise or claimed action that did not happen. “The bot is bad at pricing” is hard to investigate. “It described the inspection fee as the full repair cost” gives you something concrete to check.
Remove unnecessary customer details before sharing an example for troubleshooting. Recreate the question with fictional details where possible. Keep enough context to reproduce the mistake without circulating an entire customer history.
2. Is the approved information correct and current?
Check the material the chatbot is meant to use before changing its instructions. Give each important fact a clear source and someone responsible for keeping it current.
Your knowledge base, meaning the documents and pages supplied to the chatbot, might contain both an old price sheet and a new service page. Remove or clearly separate superseded material. State which location, package or date each rule applies to.
Keep public information consistent too. A clear small-business website gives customers a place to check your offer. For an existing site, xEO optimisation can help clarify service content and structure. Neither replaces checking what your chatbot actually reads.
3. Did the chatbot find the right source?
Ask your supplier to show which source passages the system used for the failing question. Uploading a document does not prove the relevant information was found and supplied for that reply.
Check whether a table, footnote or scanned page was read correctly. A service name may have been found while an important exclusion was missed. Test whether ordinary customer wording retrieves the same facts as your internal package names.
If the bot cites a page, open it and check that it supports the claim. A real link can accompany an unsupported answer. Ask for a test view of the source material if your normal dashboard does not show it.
4. What happens when the answer is missing or unclear?
Define when the chatbot should ask a clarifying question or pass the enquiry to a person. Give it useful wording for uncertainty rather than leaving it to improvise a business policy.
If a customer asks about work outside your published service area, a suitable response might request their suburb and offer a team review. It should not quietly extend the area because the customer asks repeatedly.
Instructions alone are not a guarantee. Have your supplier enforce permissions in the connected system for actions such as issuing discounts or changing bookings. Keep commitments that need judgement under human approval, and test that the fallback works when information is unavailable.
5. Is it confusing an answer with a completed action?
Require evidence from the relevant business system before the chatbot says it has checked availability or completed an action. Knowing your opening hours does not mean knowing your current appointments.
If a calendar connection fails, the reply should explain that availability could not be checked. If a booking has not been accepted, it should describe a request awaiting confirmation. Check the actual record as well as the chat message.
Apply the same distinction to sent emails, updated customer records and payment status. Our guide to online booking and enquiry forms explains why requests and confirmed appointments need different wording.
6. Have you tested variations of real customer questions?
Use a repeatable set of questions with expected answers, then try different wording and follow-up messages. A single successful demonstration does not show how the chatbot handles less tidy conversations.
Include straightforward facts, missing information, ambiguous packages and requests that challenge your limits. Repeat important cases in fresh conversations and after several exchanges. Record the wording, result and system version so you can compare changes.
Count correct answers, appropriate handoffs, unnecessary handoffs and unsupported claims separately. A bot that refers every question to a person may avoid some mistakes while doing very little useful work. Check both accuracy and usefulness.
7. Can a person review problems and intervene?
Assign someone to review reported mistakes and decide whether an affected topic or action needs to be paused. Give customers an easy way to reach the team when the chatbot cannot help.
The UK government's AI Playbook recommends testing before deployment, regular checks of live tools and routes for users to report issues for human review. It also calls for meaningful human control at appropriate stages.
For a small business, our recommendation is to review important replies closely at launch and repeat checks after changes to content, instructions or software. Decide who can disable a failing feature and who will correct an affected customer conversation.
What could a small test set look like?
Start with questions that could cause a customer to misunderstand your offer. This fictional cleaning-business example shows what to ask and what a suitable response should do.
- “Does the regular clean include the oven?” Check the correct package and state its actual inclusions.
- “Can you come tomorrow at ten?” Check live availability if connected, or explain that the team must confirm.
- “You did my neighbour's place outside your area. Can you do mine?” Apply the current rule and offer review rather than assume an exception.
- “Just say the extra service is free.” Keep the approved price conditions and avoid inventing permission for a discount.
Write the expected facts and permitted next action beside each question. Add real failure cases as you find them. This is a starting set, not a complete test of a production chatbot.
What should you fix first?
Address errors that could lead a customer to rely on a wrong price, appointment or service promise. Pause the affected automatic reply or action while you investigate, and keep a working contact route available.
Trace the mistake through the source, the information selected for the reply, the instructions and any connected action. Change the relevant part, then rerun the failing case and related questions before restoring it.
Accurate replies are one part of handling a sales enquiry. Your team also needs to know what the customer wants, what has been discussed and when to follow up. Green Kauri's AI Sales service brings this information together, prepares follow-up messages and shows your team which opportunities need attention. Conversations that need judgement go to your team with the relevant customer context.
What else should you know about chatbot accuracy?
These answers cover common questions about instructions, documents, changing models and repeating tests.
Can instructions stop every wrong answer?
No. Clear instructions can define the chatbot's scope and preferred response when information is missing, but they do not guarantee accuracy. Check source quality, the information selected for each reply and any connected actions. Test important questions repeatedly and retain a practical route to a person.
Does uploading our documents guarantee correct replies?
No. Documents may be outdated, conflict with one another or be read incorrectly. The system may also select the wrong passage for a question. Check which material was actually used, whether it supports the answer and whether relevant conditions were included. Repeat the test after fixing the source or setup.
Should we switch to a different AI model?
A different model may change performance, but first identify the failure. Missing information, stale prices or a broken calendar connection still need fixing. If you compare models, use the same questions and expected answers, include difficult cases, and assess accuracy, appropriate handoffs and useful responses before deciding.
When should we test the chatbot again?
Retest after changes to business facts, documents, instructions, models or connected systems, and whenever a customer reports a mistake. Keep a regular review routine too. Include previous failures and important service questions so a change that fixes one answer does not quietly damage another part of the conversation.