We measured 2,544 AI receptionist calls. It handled fewer than four in ten

Every AI receptionist vendor will tell you what their agent can do. None of them will tell you what share of your calls it will actually finish.

So here is a real answer, from a live South African deployment rather than a demo. Three months of inbound calls, May to August 2026, counted in full.

The basis matters, so it is worth one sentence: of 2,544 calls that reached the agent, 392 were abandoned by the caller during the greeting and are set aside, leaving 2,151 engaged calls as the denominator throughout. The numbers are less flattering than the marketing and more flattering than the backlash, and both are worth knowing before you sign.

Fewer than four in ten calls were finished by the AI

Of 2,151 engaged calls, the agent resolved 813 on its own. That is 37.8%.

The remaining 1,338 calls, 62.2%, either went to a person or ended without resolution.

Set that against how these products are sold. The pitch is that the AI frees your receptionist. On this evidence it freed them from roughly a third of the calls, which is a real result, and a long way from the impression most buyers have when they sign.

The handovers were fast, which matters more than the percentage

The median transferred call lasted 42 seconds and ran to nine messages.

That is a greeting, a caller saying what they need, the agent recognising this is not something it can finish, and a handover. Half of all transfers were done inside 32 to 59 seconds. Only about 2% ran past two minutes.

This cuts against the usual criticism of AI voice agents, so it deserves saying plainly: there was no mass of callers stuck in long, circular conversations before being dumped on a person. The agent recognised its limits quickly and acted on them.

So the honest reading of a 62.2% handover rate is not “the AI failed 62% of callers”. It is closer to “the AI triaged accurately and drew its boundary in under a minute”. Whether that is worth what you pay for it is a separate and fair question.

What the other 62% were actually about

Two things send a call to a person, and only one of them is a technology problem.

The first is simple preference. A caller asks for a human, and the agent obliges. No configuration fixes that, and arguably there is nothing to fix.

The second is the interesting one. A caller asks about something the agent was never set up to handle, because nobody knew people called about it. Every deployment starts from a list of the reasons a business believes its customers phone. The call record almost always shows a different list.

That gap is not an AI defect. It is a business discovering, often for the first time, what its customers actually want, and it is frequently worth more than the receptionist hours saved. A caller reason that turns up 40 times a month is either worth configuring for, worth fixing upstream, or worth routing to a person on purpose. Those are business decisions, and the agent will keep surfacing them for as long as it runs.

One configuration change moved the result by 13 points

This is the finding that should shape how you resource the project.

Partway through the measured period the agent’s configuration was updated. From that morning onward, across the next 1,387 consecutive calls, the share of calls handed to a person ran roughly 13 percentage points higher than it had under the previous setup, and it stayed there.

A higher handover rate is not automatically worse. An agent that recognises its limits sooner may be doing exactly what you want it to do. The point is the size of the swing: one change to the setup moved a headline operating metric by double digits, and nothing about the phones, the network or the callers had to change for that to happen.

Worth noting that call volume also rose over the same period, so the mix of callers may have shifted alongside the configuration. The relationship is strong but this is field data, not a controlled experiment.

The operating conclusion holds either way. An AI receptionist is not an appliance you install once. It is a configuration sitting against a business that keeps changing, and its behaviour can move significantly on a single edit. If nobody is reading the numbers, nobody will notice when it does.

What this means if you are buying

Nothing here says do not buy an AI receptionist. It says buy it with the right expectation and the right budget.

Roughly a third to a half of your calls being finished by software is a genuine result. It is not the result most buyers picture when they sign, and the gap between those two numbers is where disappointment lives.

Keep reading