AI receptionist

What AI Receptionists Cannot Do, and When to Hand Off

A practical look at the limits of an AI receptionist, the cases that should always reach a person and how to design hand-offs that never leave a caller stuck.

Call editorial team · Published · 5 min read

Marketing for automated phone answering tends to focus on what it can do. A fair guide has to spend time on the other side: what it cannot do, where it tends to fail, and how to build a line that fails gracefully. Businesses that do this well are usually the ones who find the technology useful for years, rather than the ones who switch it off after a bad week.

A receptionist at a front desk transferring a call to a colleague
Illustrative image

This article lists the limits that matter in practice and shows how to design hand-offs around them. If you are still deciding whether to use an automated receptionist, begin with What Is AI Phone Reception?. If you are ready to set one up, the companion piece is AI Receptionist Setup Checklist.

Where an AI receptionist works well

To see limits clearly, start with the strengths. An automated receptionist can answer at any hour, repeat the same correct information without tiring, take a message in an organised form and route a call by what the caller says. It is designed to handle the large, boring middle of your call volume: hours, location, how to book, what to bring, who to speak to.

Those strengths come with a condition. The system is only as reliable as the information it is given. If the source says the shop closes at six and it now closes at five, the system will say six. Good results depend on maintained content and regular listening.

The limits that matter

What it cannot do well

  • Judge urgency or safetyIt cannot reliably tell a minor worry from a medical, safety or legal emergency.
  • Make decisions with authorityRefunds, exceptions, discounts, diagnoses and contracts need a person who can be accountable.
  • Handle strong emotionAn upset or grieving caller needs empathy that a script cannot supply.
  • Know what it has not been toldIt may sound confident about an area outside its sources.
  • Cope with every voice and lineAccents, background noise, poor connections and overlapping speech can cause mishearing.
  • Verify identity on its ownA voice or a number is not proof of who is calling.

The third and fourth items catch people out. A caller who is crying does not want a cheerful menu. A caller who asks a question you never anticipated may receive a plausible answer that is wrong. Both situations argue for a quick path to a person and for sources that include an explicit "I cannot confirm that" behaviour.

Cases to hand off every time

Write a short list of situations that always go to a person, and test that each one does. The list varies by business, but most contain the following.

A hand-off that helps and one that does not

Dead endHelpful hand-off
Caller asks for a personMenu repeatsCall rings the people on duty
Nobody availableCall ends or loopsCaller can leave a message or ask for a callback
Out of hoursSilence or a long recordingOpening times, then message or callback
Emergency wordingTreated like any other callPlain instruction to use the emergency service, plus an alert
What the team receivesNothingA summary with the caller's words and contact details

Notice that the helpful column does not require a person to be available every minute. It requires that the caller always knows what will happen next. For the design of options to reach a person, see Letting Callers Reach a Person Quickly.

Being honest with callers

Trust depends on honesty about what the caller is dealing with. Callers who ask whether they are speaking to a person should be told the truth. A system that pretends to be human may work for a few calls and then cause real anger, and rules in some places require disclosure. See AI Disclosure in Calls for the reasoning.

Honesty also applies to capability. If the system cannot book, say so, and explain how the caller can request a booking. If it cannot confirm a price, say that a person will. A short, true answer is better than a long, wrong one.

Finding the limits in your own calls

The most reliable way to learn the limits is to listen. Review a small sample of calls every week, concentrating on those where the caller asked for a person, hung up early or repeated themselves. These are the calls where the system reached its edge.

A weekly review that finds the edges

  1. Pick the callsChoose calls that ended in a hand-off, a hang-up or a long exchange.
  2. Listen or readNote where the caller became frustrated or the answer was wrong.
  3. ClassifyMissing information, unclear wording, a case that should always go to a person, or a speech problem.
  4. FixAdd the fact, reword the answer or add the case to the hand-off list.
  5. RetestCall the line yourself and confirm that the problem is gone.

Expectations inside the team

Limits are also a people issue. Staff who expect the system to handle everything will be frustrated by hand-offs. Staff who expect it to fail will not trust it. Explain what it handles, what it passes on, and what the summary will look like. Make it easy to report a wrong answer, and respond visibly when someone does.

A limits checklist for launch day

  • The always-hand-off list is written and each case has been tested
  • Every hand-off ends in a person, a message or a callback
  • Emergency wording points to the emergency service
  • The system says plainly that it is automated when asked
  • Sources are dated and someone owns keeping them current
  • Staff know where to find call summaries and how to report a mistake

Handing off in Call

In Call, the AI agent answers only from the instructions and sources you give it. When nothing there covers a question, it says that it cannot confirm that on this call, and offers a person where the line can put callers through. If a caller asks whether they are speaking to a person, it says it is an AI assistant. It does not book appointments, send anything or give discounts on the call.

When a caller asks for a person, the call rings the members you have chosen at once, and the first to answer takes it. If nobody answers, the caller can leave a message. Those callers appear in a help-desk list with the request, a summary and the conversation, and staff can assign it, add a note and call back. A page shows, for each line, who would ring right now and whether business hours are open, so you can check the hand-off before callers rely on it.

Frequently asked questions

What should an AI receptionist never do?

It should not judge emergencies, give professional advice, make decisions on refunds or exceptions, or confirm identity on its own. These need a person with the right knowledge and authority.

How do I know when it has reached its limits?

Listen to calls that ended in a hang-up, a request for a person or repeated questions. Those show where it lacked information or misunderstood.

Should the system say it is an AI?

Yes, when asked, and many businesses say it at the start. Honest disclosure builds trust and may be required by law in some places.

What if nobody can answer the hand-off?

Offer a message or a callback request, and tell the caller what will happen next. Avoid dead ends such as looping menus.

Can an AI receptionist cope with strong accents or poor lines?

Performance varies. Test with realistic calls, including noise and different accents, and always provide a route to a person after repeated misunderstanding.

How often should we review the setup?

Weekly at first, then monthly. Also review after any change to hours, prices or services, because the system will repeat whatever it was last told.

Set up caller hand-off in Call

In the demo console you can try incoming calls in a simulation.

Related guides