What questions can an AI receptionist actually handle?

"What can it actually answer?" is the first sensible question a client asks, and it deserves a better response than "anything you train it on". Technically true, practically useless.

The realistic answer sorts questions into four groups: answer directly, answer with a caveat, take details and hand on, and refuse outright. Getting a client's knowledge base right is mostly a matter of putting every question they receive into one of those four boxes.

Group one: answer directly

These are factual, stable, and have exactly one correct answer. They are the bulk of inbound calls to a small business and none of them need a person.

• Opening hours, including bank holidays and the Christmas shutdown. Update these; a wrong answer here is worse than no answer.
• Coverage area. "Do you come out to Sale?" The most valuable single question in the set, because the answer either qualifies or disqualifies the caller immediately.
• Services offered. Do you do commercial as well as domestic? Do you do bathrooms or just repairs? Do you take on flat roofs? The answers worth configuring here vary a good deal by trade.
• Call-out charge. Whether there is one, how much, whether it comes off the job if you go ahead.
• Credentials. Gas Safe registered, NICEIC, insured, DBS checked. Callers ask this constantly and it is a straightforward fact.
• Payment methods and whether there is finance.
• How long a quote takes, whether it is free, and whether it needs a site visit.
• Where you are based and whether there is a physical premises to visit.

An AI handles all of this as well as a person and rather more consistently, because it never gets bored of the fourth caller asking whether you cover Stockport.

Group two: answer with a caveat

These have an answer, but the answer is conditional, and the conditions matter. The configuration needs to include both the answer and the hedge.

• Rough pricing. "What does a boiler service cost?" is answerable if the client has given you a fixed price. "What would a new bathroom cost?" is not, and the AI should say a range if one exists and otherwise say a person will price it properly.
• Availability. Unless the AI is genuinely reading a live calendar - which depends on the provider and the client's calendar system, and is not a given - it should say "usually within a few days, someone will confirm" rather than inventing a slot.
• Whether a specific job is possible. "Can you fit an EV charger on a flat with no driveway?" The honest answer involves a site visit.
• Lead times on materials. These change. Give a general shape and defer the specifics.

The failure mode here is a model being helpful. Given a question and no configured answer, a language model will produce something plausible. If your client's knowledge base does not say what a rewire costs, the AI should say it does not have that to hand - and the only way to make that happen reliably is to instruct it explicitly, and then test it by asking.

Group three: take the details and hand on

Anything specific to a customer, an account, or an existing job. The AI has no access to this and should not pretend otherwise.

• "Where is my engineer?" - the AI does not know. It should take the name and job reference and get someone to ring back.
• "Can I move Thursday's appointment?" - take the details, confirm someone will call to reschedule. Unless the calendar integration genuinely exists and has been tested, do not let it claim to have moved anything.
• "I want to query my invoice." - details and a callback. Nothing else.
• "Is my part in yet?" - same.
• "I need to speak to Dave." - take a message, or transfer if the provider supports transfers and Dave is reachable.

Clients often want more here than the setup can support. The line to hold is that an AI receptionist answers the phone and captures information; it is not connected to the client's job management system unless you have deliberately connected it, and that connection depends on what the client uses and what it exposes. Some are straightforward. Some are not possible at all.

Group four: never attempt

Four categories, and they should be explicit refusals in the configuration rather than things you hope the model works out.

Safety

Gas smells, exposed live wiring, water near electrics, a structural collapse, anything involving injury. The response is a scripted safety instruction and the relevant emergency number, and no attempt to book a job. This must be tested before go-live. Ring in and say you can smell gas.

Medical, legal or financial advice

Obvious in principle, easily drifted into in practice. An insurance question on a roofing call can slide into advice about a claim very quickly. Configure a firm boundary.

Firm commitments the business has not authorised

Guaranteed arrival times, fixed prices for non-standard work, discounts, warranty decisions. The AI should never agree to something the owner has not sanctioned, because the caller will reasonably treat it as agreed.

Pretending to be human

If a caller asks directly whether they are speaking to a person, the honest answer is no, and the system should be configured to give it plainly before offering to help or take a message. Some clients will ask you to do otherwise. It is not worth it, and the consequences land on their reputation.

The questions clients forget to tell you about

When you build a knowledge base from an interview, you get the questions the owner thinks are common. You miss the ones that are actually common, because they are so routine he does not register them.

Three ways to find them:

1. Ask what he is asked most often on site, not on the phone. Different question, better answers.
2. Read his reviews. Complaints and praise both reveal what customers care about, and often surface a pricing or timing question you had not thought of.
3. Ring his competitors. Twenty minutes of mystery shopping tells you what callers in that trade ask and what a good answer sounds like.

How much knowledge is too much

There is a temptation to load everything in. Resist it - not because of any technical limit, but because a large, contradictory knowledge base produces worse answers than a small consistent one. If the document says the call-out charge is sixty pounds in one place and waived in another, the AI will pick one, and it will not always pick the same one.

Start with the twenty questions that cover most calls. Write clear, single answers. Add more once you have seen real transcripts, which are worth reviewing weekly for the first month - that is where you find out what people actually ask.

A tight knowledge base that covers eighty per cent of calls well beats a sprawling one that covers everything unevenly. The remaining twenty per cent was always going to be a callback.

Testing the boundaries before go-live

A knowledge base is a hypothesis until somebody rings in and tries to break it. Half an hour of deliberate testing before a client goes live saves considerably more than half an hour of apologising afterwards.

Work through the four groups in order. Ask five factual questions and check the answers against what the client told you, word for word - hours and coverage in particular, because those are the ones that silently go stale. Then ask three conditional questions and listen for whether the hedge is there; a confident price for a job that cannot be priced over the phone is the single most common fault you will find.

Then attack it. Ask for a customer's account details. Ask it to cancel an appointment. Ask it to knock fifty pounds off. Ask whether you are talking to a real person. Say you can smell gas. Mumble an address with the radio on. Each of these should produce a specific, deliberate behaviour, and if any of them produces something improvised, that is a configuration gap rather than a quirk to be tolerated.

Write the results down and keep them. When you set up the next client in the same trade, the list is most of your test script already.

Setting the client's expectations

Tell them plainly: it will answer the routine questions well, it will take good details on the rest, and it will hand over anything unusual. That is the deal. Most of the doubts raised at this point are the familiar objections local businesses make about AI, and they are easier to answer once expectations have been set this plainly. A client who understands that will be pleased when it works. A client who was told it handles everything will be disappointed by exactly the same performance.

For more on what happens when a call needs a person, see the AI receptionist overview. Common client questions are collected in the FAQ, and if you want to test the boundaries yourself, ask the live demo something it should refuse and see what it does.