What can go wrong with an AI receptionist — and how to find out safely

The real risks are mishandled complex calls, mis-heard names and addresses, and a system that answers confidently when it should have taken a message. All three are testable before you commit. The safe way to find out is conditional forwarding, pass and fail criteria you set in advance, and calling the number yourself before any customer does.

• Conditional forwarding means the AI only ever handles calls you would otherwise have missed.
• Decide your pass/fail thresholds on day 0 — deciding them on day 14 is just rationalising.
• Read transcripts, not only summaries. A summary cannot tell you what it invented.
• Know the deactivation code for your carrier before you turn forwarding on.
• The free trial runs 14 days or 50 minutes, whichever comes first — plan around the minutes.

What actually goes wrong with an AI receptionist?

Four things, in rough order of how often they bite. None of them are secret, and any vendor who tells you their system has no failure modes is telling you something useful about the vendor. Knowing these in advance is what lets you design a pilot that would actually detect them.

• Complex or emotional calls. A distressed caller, a complaint, or a genuinely unusual situation is where a human is clearly better. The AI should recognise it and take a message or transfer — not try to resolve it.
• Mis-heard details. Unusual surnames, street names, and spelled-out email addresses are the common casualties, and a bad connection makes it worse. This is why callback numbers matter more than any other captured field.
• Answering what it should not. The expensive failure is a confident answer to a pricing or policy question you never authorised. Constrain what it may state, and test that boundary deliberately.
• Gaps in what it knows. If your business rules live only in your head, the AI cannot apply them. Undocumented knowledge is the single biggest predictor of a disappointing pilot.

Will customers mind that it's AI?

Most callers care far more about reaching someone quickly than about who picks up. The reaction that actually causes damage is not recognising it as AI — it is being trapped with something that cannot help and offers no way through to a person.

So the thing to test is not whether it sounds human. It is whether a caller who needs a person can get one. Make sure there is a route to a human or a callback, and check during your pilot that it triggers when it should rather than looping.

What happens on a call it can't handle?

It should take a detailed message, capture contact details, and optionally transfer — a graceful handover rather than a dead end. What you are checking is that the fallback is a real path and not a polite way of ending the call.

Test this deliberately by asking it something outside its knowledge and seeing what happens. The answer you want is an honest "I'll take your details and have someone call you back." The answer that should worry you is a plausible-sounding invention. This is question five in how to choose an AI receptionist.

How do you test one without risking your business?

Run a bounded pilot where the AI only ever gets calls you would have missed anyway. That single design decision removes almost all the downside, because the comparison is against a voicemail nobody listens to rather than against you answering the phone well.

1. Get your trial number and do not forward anything yet.
2. Call it yourself twice — once as a straightforward customer, once being deliberately awkward. Use an unusual name and a difficult address.
3. Ask it something it should refuse to answer, like a firm quote on a job that varies. Check that it declines rather than invents.
4. Turn on conditional forwarding so your phone rings first and only unanswered calls reach it.
5. Tell one member of staff what is happening, so somebody is watching for odd customer reactions.
6. Read every transcript for the first few days, not just the summaries.
7. Compare your carrier's missed-call log against the receptionist's call list at the end.

What should your pass and fail criteria be?

Write them down before you start, with numbers you choose. Judging a pilot after the fact is how people talk themselves into keeping something that did not work, or dropping something that did. Five criteria are enough for a first pilot.

Fill in your own thresholds in the middle column before day one, and keep a simple daily tally against them: calls answered, calls that still needed you, bookings or leads captured, and anything the receptionist got wrong. Two weeks of that beats any amount of deliberation. If the numbers do not clear the bar you set, the payback arithmetic will tell you whether it was close.

How do you get out if it isn't working?

You dial your carrier's deactivation code and forwarding stops immediately. Your line goes back to behaving exactly as it did before, with the same number and the same voicemail. Nothing needs unwinding on the receptionist's side first, and nothing about your number was ever transferred.

This is the strongest practical reason never to port your number during a trial. Forwarding is reversible in seconds; porting is another multi-day process. The codes for each carrier are in keeping your business phone number — note yours down before you switch forwarding on, not after.

Related questions

Can it damage relationships with existing customers?

The risk is real for high-touch relationships where clients expect you personally. Mitigate it by starting with conditional forwarding and after-hours coverage, so regulars still reach you directly during the day and the AI only catches what would have gone unanswered.

What if it books an appointment I can't honour?

Constrain availability at the calendar level rather than trusting the conversation. Check bookings the same day for the first week — that is when configuration mistakes surface, and a same-day catch is a phone call rather than a missed appointment.

Is my call data secure?

Ask any vendor where recordings and transcripts are stored, how long they are kept, and who can access them. If you handle health, legal, or financial information, confirm your own regulatory obligations first — that duty stays with you regardless of the tool.

How long should a pilot run?

Two weeks covering at least one full weekend, which is where after-hours coverage proves itself. Shorter than that and normal week-to-week variation swamps the signal you are looking for.

What to do next

• Write down your five pass/fail thresholds before you start, using the table above.
• Work out whether the numbers justify it at all: is an AI receptionist worth it?
• Set up conditional forwarding using your carrier's code, and write down the code that turns it off.