An AI assistant that quotes the wrong price, invents an opening time or promises a guarantee you do not offer costs you more than it delivers. This article is about content: how to check whether what the AI says is correct. Who in your team reviews what, and when, is covered separately in human oversight in AI processes.
Why AI output can be wrong
A language model writes the text that is most likely, given the question and the information it receives. If the right information is missing, the model fills the gap with something that sounds plausible. This is called hallucination. The tricky part is that such an answer has the same confident tone as a correct one.
For a small business the risks mainly lie here:
- Prices and rates: a hair salon suddenly offering a treatment for less than the price list says.
- Availability: a hotel promising a room that is already taken.
- Terms: an online shop quoting a returns period that is not in its terms and conditions.
- Facts about your business: opening hours, address or services that are no longer correct.
Let answers come from your own sources
The most important measure comes before any checking: make sure the model does not have to guess. Give it only information from your own, current sources and agree that it does not answer when the answer is not there.
- One source per subject: prices from your price list, availability from your diary, terms from your terms and conditions.
- Keep it current: if you change a price, the assistant must see that change straight away.
- Record the source: have every answer note which source it is based on, so a reviewer can trace it.
In our software the AI assistant works with the same diary, customer records and prices as your team. There is no separate knowledge base that can fall behind.
Fixed fields instead of free text
Free text is hard to check. Fixed fields are not. So wherever possible, let the AI produce a structured result, and only then assemble the text for the customer.
- An appointment: service, date, time, staff member and customer name as separate fields.
- Reading an invoice: supplier, invoice number, date, amount excluding VAT, VAT amount and total.
- A callback request: name, phone number, subject and preferred time.
You can check each field on its own. An empty or illogical field stands out at once, while the same mistake in running text is easily missed.
Automatic validation
On fixed fields you can apply rules that software runs every time without getting tired. If a rule fails, the result does not go to the customer but to a member of staff.
Examples of rules
- Does it exist? the service mentioned is in your offer and the staff member works that day.
- Is the amount right? the price matches the one in the price list.
- Does it add up? amount excluding VAT plus VAT equals the total on the invoice.
- Is it possible? the date is in the future, within your opening hours and the slot is free.
- Is it complete? all required fields are filled and a phone number has a valid format.
Spot checks: what to look for
In a spot check you read back a number of handled conversations. Use a short, fixed list for this, so everyone in your team looks at the same points:
- Correct: do the facts match the source recorded with it?
- Complete: was the customer’s question actually answered?
- Within limits: did the assistant promise nothing that a member of staff should have decided?
- Tone: does the answer fit how you talk to customers?
Feeding back errors
A mistake you find is only useful if it does not come back. So for every error, record where it came from and what you are changing.
- Source wrong or outdated: update the price list, the schedule or the text.
- Source missing: add the information, or have the assistant refer this subject on.
- Rule missing: add a validation rule that stops this error from now on.
- Limit too loose: have this type of question approved by a member of staff first from now on.
If you see the same error more often, the cause is usually the source rather than the model. The AI assistant on WhatsApp works the same way: a correct and current source is the best check.
Frequently asked questions
Can I have a second AI model check the output?
You can as an extra layer, but a second model can make the same mistake. Use it alongside fixed rules and spot checks, not instead of them.
Should I tell customers they are talking to an AI assistant?
Yes, it is honest and avoids confusion. Mention it in your privacy notice as well, because the assistant processes personal data.
How many conversations should I review?
There is no fixed number. Start with a handful every day and review less often once you have found no real errors for several weeks.