The complaint is about an hour spent standing in the rain. The customer had booked his car in for its MOT and planned to wait, because the garage's website receptionist told him there was a waiting area with coffee. There was, until the garage converted it into a second bay last spring. On the day, he was told to come back in an hour.
Nobody at the garage saw that conversation until the complaint arrived. That is what makes wrong answers from an assistant so expensive: the customer acts on them immediately, and you find out last.
AI receptionist accuracy is often discussed as if the goal were to answer everything. It is not. The goal is to answer correctly when the facts are there and to say "I'm not sure" when they are not, then make sure a person picks the question up.
The arithmetic of a wrong answer
These numbers are an illustration. Use your own.
Say your receptionist handles 200 conversations a month. In 10 of them, the visitor asks something your content does not cover.
If the receptionist guesses, perhaps half the guesses are right by luck. The other 5 are wrong. Say 2 of those customers act on the wrong answer: they turn up at the wrong time, expect a service you do not offer, or budget around a price that is not real. Each one costs you a difficult conversation, maybe a refund or a free extra, and possibly a review.
If the receptionist says "I'm not sure" instead, all 10 visitors get an honest reply and a route to a person. Some will not bother leaving details. You lose a few enquiries you might have won. But nobody has been misled, and you have 10 new questions to answer in your content.
Across a year, the second path produces a better receptionist. The first produces a slowly growing pile of small betrayals.
Why a good receptionist refuses to guess
The reason receptionists make things up is that language models are very good at producing plausible sentences. Plausible is not the same as true. Ask one about a waiting area and it will happily describe a pleasant one.
The protection is grounding: the receptionist answers only from content you have supplied, and when that content has nothing relevant, it says so rather than filling the gap. SpideyChat works this way. It answers from your website, documents and Q&A pairs, and when there is no relevant content it tells the visitor, captures the enquiry and records the question as a knowledge gap. How grounding keeps AI assistants accurate explains the mechanics, and the glossary covers the terms.
Grounding is not a guarantee on its own, though. If your website still mentions the waiting area, a grounded receptionist will mention it too. AI receptionist accuracy depends on the content being right.
Two ways to say you don't know
Being uncertain is the easy part. Being uncertain usefully is where most setups fall down. Compare these.
Visitor: Can I bring my dog to the class?
Assistant: I'm sorry, I don't have that information. Please contact us for more details.
That is honest, and it is almost useless. The visitor has to find the contact page, write the question again and wait. Many will not.
Visitor: Can I bring my dog to the class?
Assistant: I don't have anything on dogs at classes, so I'd rather not guess. I can pass the question straight to the studio team. If you leave your name and email, they'll reply, usually within a working day. Classes themselves run Monday to Saturday if you'd like the timetable in the meantime.
Same honesty. But the second version keeps the conversation alive, captures the enquiry and still gives the visitor something useful.
What a good uncertain reply includes
- A plain admission. "I don't have information on that" or "I'd rather not guess". No waffle, no apology spiral.
- What it can confirm, if anything relevant. The timetable, the address, the related fact it does know.
- A route to a person. Taking details, opening a ticket, or handing over to someone on duty if a team member is available.
- A realistic expectation. When the visitor will hear back, in words that match what actually happens.
Leave out any of these and the reply feels like a dead end. The fourth point matters more than people think. "Someone will be in touch shortly" on a Friday evening, followed by silence until Tuesday, turns an honest answer into a broken promise.
For questions that should never be answered at all, such as medical, legal or financial advice, the same structure applies with one change: the receptionist says it cannot advise, and points urgent matters to a qualified person or the emergency services.
Turning each gap into an answer
Every "I'm not sure" is a free piece of research into what customers want to know. The routine that makes use of it is short:
- Once a week, open the list of unanswered questions.
- Group duplicates. Five ways of asking about dogs are one gap.
- For the top few, write the answer as a Q&A pair, or fix the page on your website that should have said it.
- Check whether any gap is really a stale fact, like the waiting area, and correct it everywhere it appears.
- Ask the receptionist the original questions again to confirm it now answers them.
Some gaps deserve a proper page rather than a Q&A pair, because human visitors are wondering the same thing. Turning unanswered questions into website pages covers when that is worth doing.
What to do this week
Ask your receptionist five questions you know it cannot answer: something about a service you do not offer, a policy you have never written down, a local detail. Read how it responds. If it guesses, that is the first problem to fix. If it says "I don't know" and stops there, rewrite its fallback so it includes a route to a person and a realistic reply time.
Then go through your content for anything that has changed in the last year: rooms, staff, prices, hours. Stale facts are behind more wrong answers than missing ones. If your receptionist keeps saying "I'm not sure" about something it should know, what to do when your agent keeps saying I don't know is the troubleshooting guide.