What happens when an AI receptionist can't help?

Every AI receptionist reaches the edge of what it can do. A caller asks something it was never set up to answer, gets upset, or simply wants a person. What happens in the next thirty seconds decides whether the caller remembers the call as handled or as a run-around.

There is good evidence that this is where customer worry sits. A Gartner survey of 5,728 customers, conducted in December 2023 and published in July 2024, found that 64% would prefer companies didn’t use AI for customer service. The top concern wasn’t wrong answers or job losses. It was that AI would make it harder to reach a person. In Gartner’s words, many customers fear AI “will simply become another obstacle between them and an agent.”

That’s the problem a good handover solves. This article covers how an AI receptionist should recognise that it can’t help, the four ways it can pass the call on, what the caller should hear, and how to test it before real customers find the gaps.

Why the handover matters more than the happy path

When an AI receptionist books an appointment smoothly, the caller hangs up and thinks about the appointment. When it fails to help and then fails to pass them on, the caller hangs up and thinks about your business.

A poor handover usually looks like one of these:

• The loop. The agent keeps rephrasing the same question because it hasn’t been told it’s allowed to give up.
• The dead end. The agent says it can’t help and offers nothing else.
• The black hole. The agent transfers the call to an extension nobody answers, or takes a message nobody reads.
• The restart. A person finally picks up and asks the caller to explain everything again.

Each of these is a design gap, not a technology limitation. The handover is something someone has to build on purpose.

Recognising when to hand over

The first job is noticing early. Escalation triggers should be specific, written down and tested, not left for the agent to judge from vague instructions.

The caller asks for a person

This is the simplest trigger and the one most often handled badly. If a caller says “can I speak to someone”, the agent should not try to talk them out of it or ask three more questions first. It can reasonably ask what the call is about so it can route it properly, but the request itself should be honoured on the first ask.

The request is outside scope

Every agent has a defined job. Anything outside it should be passed on rather than improvised. For a trades business that might be a dispute about an invoice; for a clinic, a question about test results; for an accountant, anything that sounds like advice. The agent doesn’t need to understand the request in full. It needs to recognise that it isn’t one of the things it handles.

The agent isn’t permitted to act

Some requests are in scope to talk about but not to action without a person: refunds, discounts, changes to a contract, cancelling a large job. We covered how to decide which actions need approval in should an AI agent ask before it acts?. On the phone, the same principle applies. The agent can capture the request and pass it to someone who can approve it.

The conversation isn’t working

Repeated misunderstanding is a trigger in its own right. If the agent has failed to understand a caller twice, a third attempt rarely helps. Background noise, a strong accent the system handles poorly, or a caller who is simply struggling with the format are all reasons to move on to a person or a message.

The caller is frustrated or distressed

Frustration should be acted on at the first clear signal, not the third. Distress is different again. A caller who is upset, unwell or dealing with something difficult needs a person, or at minimum a clear and calm path to one, and the agent should not try to finish its script first.

Emergencies

An AI receptionist is not an emergency service and should never be designed to sound like one. If a caller describes a situation that could be an emergency, the agent should tell them plainly to hang up and call Triple Zero (000). The federal government’s Triple Zero page is clear that 000 is for police, fire or ambulance in an emergency, that 112 also works from a mobile, and that people who are deaf or have a hearing or speech difficulty can call 106 by TTY.

What counts as urgent for your business is a separate question. A burst pipe for a plumber or a client locked out of a property for a property manager is urgent without being a 000 matter, and those calls should follow your own urgent-escalation path.

Relay calls

Some callers reach you through the National Relay Service, where a relay officer joins the call and passes on what each person says. The NRS guide for businesses explains that relay calls take longer, that you should speak directly to the caller rather than the officer, cover one topic at a time, and wait for the caller to respond.

Those are exactly the conditions that can trip up a voice agent tuned for fast, back-and-forth conversation. Unless you have tested your agent with relay calls and are confident it handles the pacing, treat a recognised relay call as a reason to transfer to a person or take a careful message.

Four ways to hand over

Once the agent knows it can’t help, it has four options. Most businesses need more than one, depending on the time of day and who is available.

Warm transfer

The agent connects the caller to a person and passes on what it has learned before the caller is put through. Depending on the phone system, that might be a short spoken summary to the staff member, a note that appears on screen, or both.

This is the best experience when someone is available, because the caller doesn’t have to start again. It also takes the most setup: the phone system has to support it, and the summary has to be short and accurate enough that the person picking up trusts it.

Cold transfer

The agent puts the call through without passing on any context. It’s simpler, and sometimes the only option. The cost is the restart problem. If you use cold transfers, at least have the agent log a summary to your CRM or inbox so the person who answers can look it up quickly.

Taking a message

When nobody is available, a well-taken message is often better than a transfer that rings out. A good message captures:

• the caller’s name and the best number to reach them, read back to confirm
• what the call is about, in a sentence or two
• how urgent it is, in the caller’s own words
• the best time to call back
• anything the caller has already been told

The message then needs to go somewhere a person will see it promptly. That might be a shared inbox, a text to whoever is on call, or a task in your CRM. We go into the mechanics in how AI agents integrate with a CRM.

Booking a call-back

A step up from a message: the agent offers the caller a specific time for someone to ring back and books it into a real calendar. This works well when the agent can see staff availability, and badly when it can’t. A call-back time that nobody has committed to is just a message with a false promise attached.

What the caller should hear

The words the agent uses at the handover matter as much as the mechanics. A few principles hold up across most businesses.

Say what’s happening. “I’ll put you through to Sarah in our accounts team now” is better than silence or hold music with no explanation.

Be honest about timing. If the call is going to a message, say so. If someone will call back today, say today only if that’s true. Callers forgive “someone will call you back tomorrow morning” far more readily than a vague promise that isn’t kept.

Don’t make them repeat themselves. If the agent has collected details, the person who picks up should already have them, and the caller should be told so: “I’ve passed on what you’ve told me.”

Give a way out. If a transfer doesn’t connect, the caller should hear what happens next, not be dropped or sent back to the start.

Here is a simple example of a handover when nobody is free:

“I’m not able to help with that one, but I can make sure the right person gets back to you. Everyone’s on calls at the moment, so I’ll take a message. Can I confirm the best number to reach you?”

And when a transfer is possible:

“That’s something our office manager handles. I’ll put you through now and let her know what it’s about, so you won’t need to explain it again.”

Both are plain, honest and specific. Neither pretends the agent is something it isn’t. The Australian Government’s guidance for AI adoption makes the same general point: people should know when they’re interacting with AI, and organisations should set up channels for people to raise problems or question what the AI has done.

When nobody answers

The handover plan has to work at the worst moment, not the best. That means writing down a fallback chain for each time of day.

A typical chain during business hours might look like:

1. Warm transfer to the relevant person or team.
2. If unanswered after a set number of rings, try a second person or a group.
3. If still unanswered, return to the agent, which tells the caller and takes a message or books a call-back.

After hours, the chain is usually shorter:

1. Recognise whether the call is urgent by your business’s own definition.
2. If it is, follow the urgent path: a text or call to whoever is on call, with the caller told what will happen and when.
3. If it isn’t, take a message or book a call-back for the next business day, and say when that is.

Two rules make this work. First, the agent should never transfer to a number that could ring out without a fallback behind it. Second, someone in the business needs to own the messages. If nobody is responsible for working through them each morning, the system fails quietly and the agent gets the blame.

Public holidays deserve their own line in the plan. They vary by state and territory, and the agent should know when your office is closed so it doesn’t promise a call-back on a day nobody is working.

Privacy when the agent takes a message

A message is personal information: a name, a number and usually a reason for calling. How much of that your business has to manage under privacy law depends on who you are, so this section is general.

At the time of writing (October 2026), the Privacy Act 1988 and the Australian Privacy Principles apply to many businesses, but not all. The OAIC explains that a small business with an annual turnover of $3 million or less is generally not covered, with exceptions including health service providers and businesses that trade in personal information. Plenty of small businesses choose to follow the principles anyway, because callers expect it.

For businesses that are covered, APP 5 requires reasonable steps to let people know certain things about the collection of their personal information, at or before the time it is collected or as soon as practicable afterwards. The OAIC’s APP 5 guidelines note that for telephone collection, this can include explaining the relevant matters at the start of the call, possibly with a script or automated message, or giving the information afterwards, for example by pointing to a notice on your website.

In practical terms, for an AI receptionist taking messages:

• Collect only what the person calling back actually needs.
• Make sure messages go to a place with sensible access controls, not a personal inbox that several people log into.
• Know where call recordings, transcripts and messages are stored, including whether any processing happens overseas, and how long they are kept.
• Have a privacy notice the agent can point to.

Separate changes to the Privacy Act covering automated decisions commence on 10 December 2026, and recording rules for calls vary between states and territories. Both are worth checking for your situation. Our article on whether business data is safe with AI agents covers storage, access and retention in more detail.

This is general information, not legal advice. Get advice on your own circumstances, particularly if you work in health, finance or another regulated industry.

Test the handover before your customers do

Most voice agents are tested on the calls they’re meant to handle. Fewer are tested on the calls they’re meant to pass on. Before going live, and again after any change to your phone system or staff roster, make test calls that cover:

• asking for a person straight away
• asking for a person halfway through a booking
• a request clearly outside scope
• a request in scope but needing approval, such as a refund
• a caller who mumbles, talks over the agent or has background noise
• a frustrated caller who says so directly
• a caller describing something urgent for your business
• a transfer during business hours when nobody answers
• an after-hours call and a public holiday call
• a message, followed by checking it arrived where it should, with the right details

For each one, check what the caller heard, where the call or message ended up, and whether the person who picked up had the context they needed. Then listen to a sample of real handovers each week for the first month or two. The gaps that matter usually show up quickly.

The short version

An AI receptionist will always meet calls it can’t handle. The ones that work well recognise those calls early, pass them on quickly, and tell the caller honestly what is happening.

Design the handover deliberately: the triggers, the transfer options, the words the caller hears, and the fallback when nobody answers. Make sure someone in the business owns the messages. Then test the awkward calls, not just the easy ones.

Get this right and callers experience one conversation that happened to involve two participants. Get it wrong and the agent becomes the obstacle customers were worried about in the first place.

If you’d like to talk through how a handover would work with your phone system and your team, the 3xRMD Voice page covers our approach. For the broader picture of where voice agents fit, see is AI voice suitable for Australian businesses?