What makes people hang up on AI receptionists? We analyzed 450,000 calls

Every AI receptionist gets hung up on. Some far more than others. We analyzed 450,702 inbound calls across 503 receptionists (each with at least 200 real, non-test calls) to find which of the setup choices an owner controls actually change that.

Two kinds of hang-up

A hang-up here is one of two things:

1. A short call: under 15 seconds. The caller connects and drops almost immediately.
2. A no-interaction call: the caller never said a word. The AI talked, the caller stayed silent, then left.

These overlap heavily (about 8 in 10 short calls are also silent), but they aren't identical, so we ran the entire analysis twice, once against each definition, and only kept findings that held up under both.

Most of the gap isn't about your settings

Take two businesses with very different hang-up rates, one low and one high. How much of that gap comes down to how they set up their AI, the greeting, the voice, the training? According to our analysis, only about an eighth of it. Roughly 13%.

The other seven-eighths comes from things those settings don't touch.

The likeliest driver is the calls themselves, not the AI. These hang-ups are mostly silent drops in the first few seconds, before the caller has reacted to anything specific, which points to junk calls rather than a real customer.

Your raw hang-up rate grades your phone number's reputation more than your receptionist. Judge the AI on the calls where someone actually speaks.

The eighth you do control is free to change. Here is what moved, ranked by size and consistency of effect.

What helps

1. Open with the business, not the owner's absence.

The biggest controllable harm we found was greetings that lead with the person being unavailable, associated with roughly 50% higher hang-up odds. Two examples (paraphrased) from accounts with very high hang-up rates:

"Hi there, [owner] has asked that I take your call while he's away."

"[Owner] is currently with clients, but I'm here to make sure you get what you need."

At first, it may sound like a good idea to tell people that the only reason you're using AI is because you're busy. But when the first thing a caller hears is that the person they wanted is gone, some decide to try again later or call elsewhere. The lowest-hang-up greetings simply skip it: "Thank you for calling [business]. How can I help you today?"

2. End the greeting with a question, not a statement.

Greetings that ask nothing, just "Thank you for calling [business]," hang up about 15% more than ones that end with a question. The likely reason is silence: end on a statement and the caller is left guessing whether to speak, and some just leave.

"Thank you for calling [business]. How can I help you today?"

Which question you ask barely matters, though. Open-ended ("How can I help you?") and specific ("Are you an existing client?") performed the same, so ask something and don't agonize over the form.

3. Say it's an AI.

The common worry is that admitting it's AI answering the phone will make people recoil. We found the opposite to be true. Disclosing it's an AI is associated with about 20% lower hang-ups. A plausible reason is that people drop calls out of uncertainty ("is this a scam or a telemarketer?"), and naming what they've reached resolves it quickly.

"Thank you for calling [business]. I am an AI assistant. How can I help you today?"

4. Tell people the call is being recorded.

Greetings that mention the call is recorded were associated with about 30% lower hang-ups. Two plausible reasons: scammers don't announce recordings, so the line reads as legitimacy; and "this call is recorded" implies a human will review it, which signals the call actually matters. If you already use call recording software , a simple line in the greeting that says so is worth keeping.

"Thank you for calling [business]. Please note this call may be recorded for quality and training purposes. How can I help you today?"

5. Give your receptionist skills.

Agents given real skills, the Upfirst features that let them send texts, transfer calls, and book appointments, plus training on the business's information, had the lowest hang-up rates of all, associated with roughly 30-40% lower odds. Take this one with a grain of salt. It probably measures who sets up carefully as much as the setup itself: a diligent owner does everything well, and we can't separate "more capable agent" from "more on-the-ball business." The direction is consistent, and an agent that can act beats one that can't, so do it, just don't bank on the full effect.

What doesn't matter

• Voice gender. The choice between male and female AI voices makes no measurable difference once everything else is accounted for, under both definitions. (For what it's worth, 83% of these businesses chose a female voice.)
• Greeting length. A short greeting and a long one hang up at nearly the same rate. Content matters; length doesn't.
• The style of your opening question. Open-ended, yes/no, or a menu of options, none clearly beat the others. Asking something matters (see above); which kind doesn't.

Method, for the curious

We took every agent with at least 200 completed, non-test inbound calls (503 of them, 450,702 calls in total) and reduced each one to two counts: short calls under 15 seconds, and silent calls where the caller never spoke.

Each greeting was labeled by a language model that read it and judged what it actually does: how it opens (open-ended question, yes/no, menu, or no question at all), and whether it discloses the AI, mentions recording, says the owner is unavailable, or lists what it can do. To those we added voice gender, greeting length, speech rate, and whether the agent had skills and training configured.

We then fit an aggregated binomial logistic regression: each agent contributes its hang-ups out of its total calls, so high-volume agents carry proportionally more weight, and we report every lever as an odds ratio (how much it multiplies the odds that a given call is a hang-up). We fit the model separately for each of the two hang-up definitions and kept a lever only if it (a) pointed the same way under both and (b) survived adding log(call volume) as a covariate, which absorbs the "busy, established number" effect. The "one-eighth" figure is a separate, call-weighted linear model of the hang-up rate on the same predictors: it explains about 13% of the variance between businesses (R squared about 0.13).

Two cautions for anyone reading the coefficients. First, this is observational, so these are associations, not causes; the clearest unresolved confound is the skills finding, where conscientious owners and capable agents travel together and we can't fully prise them apart. Second, with 450,702 calls every coefficient lands at p < 0.001, so the p-values tell you nothing, we ranked on effect size and on whether a result held across both definitions instead.

The result we most want to nail down is the recording line, the strongest lever we can't yet call causal. The test is simple and it's next: take numbers that added "this call is recorded" and compare the same line, same traffic, before and after.