Does AI phone calling actually work? What happened when we let it dial, and what the evidence says.

Every vendor demo of an AI phone agent shows the same thing: a call is answered, a natural conversation follows, a result appears. What the demo never shows is the dial tone. Most calls to most people are not answered by a person, and an AI agent does not change that.

We built AI voice screening into our own recruitment platform and let it call real applicants over a ten week period, with consent gathered first and the agent stating plainly that it was an AI. Because every call is recorded in the same ledger as everything else the platform does, we can say exactly where those calls went. We have since added the only controlled evidence we know of, from a 2026 field experiment, and the rules and prices a UK buyer needs before making the first call.

When a person did answer, the conversation held. The typical screening call ran a little over three minutes, long enough to cover the questions a recruiter would ask in a first call, and nineteen in twenty of those conversations produced a complete, scored assessment the recruiter could read before deciding whether to book an interview.

That is the honest case for the technology. It is not that AI answers more calls. It is that when a call is answered, the AI holds a consistent, recorded, fair conversation at any hour, and writes it up in a form a person can act on.

Our own data says what happens when an AI dials. It cannot say whether an AI conversation is as good as a human one, because nobody was randomised. A working paper posted to arXiv in July 2026 by Brian Jabarian of Carnegie Mellon and Luca Henkel of Erasmus University Rotterdam does exactly that. In a natural field experiment with 70,000 applicants for real jobs at a recruitment-process outsourcer, applicants were randomly assigned to be interviewed by human recruiters or by AI voice agents, after which human recruiters evaluated the interviews and made the hiring decisions.

Applicants interviewed by the AI were 12 per cent more likely to receive a job offer: an offer rate of 9.73 per cent against 8.70 per cent for the human arm, a difference of 1.03 percentage points. Job starts were 18 per cent higher (6.71 per cent against 5.65 per cent), retention at one month was 18 per cent higher, and the authors found no decline in the productivity of the workers hired, measured on handle time, quality scores and customer satisfaction. Where applicants were allowed to choose, 78 per cent chose the AI interviewer. The mechanism the authors propose is controlled variance: the AI interviews were more structured and consistent while remaining responsive, and collected more hiring-relevant information.

The caveats are the ones a careful buyer would expect. The setting was entry-level customer-service roles in the Philippines at a single outsourcer, over three months in 2025; applicants rated the AI interviews significantly less natural; recruiters knew which arm an applicant was in; and the paper is not yet peer reviewed. It is nonetheless the largest controlled comparison published, and it points the same way as our production data: the advantage of AI voice is consistency and coverage when a conversation happens, not reach.

Working paper, v1 30 July 2026, v2 11 September 2026, not peer reviewed. 67,056 applicants randomised across three arms (AI 40,103; human 13,557; applicant’s choice 13,396). Setting: entry-level customer-service roles, one recruitment-process outsourcer, the Philippines, March to June 2025.

If seven calls in eight do not reach a person, the value of the system is decided by what it does with the seven, not the one. Ours sends a text message within the same minute, with a link to complete the same screening by chat, and records that the voicemail was reached so nobody redials the same person an hour later.

The same applies to the switchboard case. When the agent hears a menu or a receptionist, it ends the call and the text goes out instead. A voice agent without a text fallback is a very expensive way to reach voicemail.

The retry rule matters too. Calling straight back after a voicemail looks like harassment and is treated as such. Retries are spaced by hours, capped in number, and stop the moment the person engages on any channel. Ofcom’s persistent misuse policy treats repeat calls to the same number within 72 hours of a first instance of misuse as an aggravating factor, which is a good rule to adopt even where the calls are not misuse.

Three operational choices move the one in eight more than any choice of model. The first is the calling window: an applicant who applied at 11pm is more likely to answer at 9am than at 3pm, and a number that rings after 8pm looks like a scam. The second is the number the call comes from. A withheld or foreign number is screened by default; a UK mobile or local number showing the caller identity is not, and PECR requires it to be displayed anyway. UK numbers from Twilio are $2.50 to $3.50 a month. The third is a text a few minutes before the call saying who will call, why and when, so the person can reply "call me at 5" instead of ignoring it. The notice does more work than the call.

Three sets of rules apply, and they apply differently depending on what the call is for. The first is the Privacy and Electronic Communications Regulations 2003. Regulation 19 says a person shall neither transmit nor instigate communications "comprising recorded matter for direct marketing purposes by means of an automated calling or communication system" unless the person called has previously notified the caller that they consent, and the caller must not block their number and must give an identity and contact details. The ICO’s guide spells out that general marketing consent, or consent for live calls, is not enough; it must specifically cover automated calls, and every automated call must include the caller’s name and a contact address or freephone number.

Regulation 21 governs live marketing calls. You must not make an unsolicited live marketing call to a number listed on the Telephone Preference Service unless that person has specifically consented, and a number is protected 28 days after it is listed. The TPS itself says organisations are required by law to screen their data against the register at least every 28 days, and registration is free. The ICO enforces this: in March 2025 it fined AFK Letters Co Ltd £90,000 for 95,277 marketing calls made between January and September 2023 without evidence of consent, following complaints to the ICO and the TPS.

Whether an AI call is a marketing call is the question that decides which rules bite. The ICO’s direct marketing guidance treats messages that "confirm or remind them about appointments", check contact details or handle an administrative matter as service messages, not direct marketing, provided they contain nothing promotional. A screening call to someone who has just applied for a job, or a reminder call to someone who has booked, sits on that side of the line. A call to someone who has never asked to hear from you is marketing, and if a machine is talking, regulation 19 applies in full. Data protection law applies either way, which is why the agent states it is an AI, consent is recorded, and an opt-out on any channel is honoured on all of them.

The third rule set is Ofcom’s statement of policy on persistent misuse, in force since 1 March 2017, which covers abandoned and silent calls from automated calling systems, with penalties of up to £2 million. If a call is abandoned, the recipient should hear an information message within two seconds that identifies the caller, explains that they tried to make contact, gives a basic-rate number to decline further calls, and contains no marketing; records should be kept for at least six months. The 2016 policy dropped the fixed 3 per cent abandoned call rate from its prioritisation criteria but still uses the rate as a factor. An AI agent that speaks the moment the call connects avoids the silent-call problem by design.

Read at source in September 2026. This is a summary for buyers, not legal advice; the ICO’s guide and the regulations themselves are short and worth reading in full.

Voice AI is priced per minute of call, in layers: the platform that runs the conversation, the language model, the speech synthesis, and the telephone line. Each vendor bundles the layers differently, so the table records what each page publishes, read in September 2026. All four publish in US dollars.

Per-minute prices as published in US dollars on each vendor’s own pricing page in September 2026. A full stack for a UK call is typically platform plus model plus voice plus telephony, so add the layers before comparing.

Take a hypothetical business making 1,000 outbound first-contact calls a month, with consent in place, and apply the shares from our production data. About 125 calls reach a person, about 500 reach voicemail, about 250 never connect, and the rest hit a switchboard or are rejected before dialling. If the conversations that happen run a little over three minutes, call it three and a half, that is about 440 minutes of real conversation. Voicemail and switchboard detection happens in the first seconds, so those calls add perhaps 250 to 300 minutes of connected time at most. Around 700 to 750 billed minutes a month, in total.

Price it on Retell’s published layers with a mid-priced model: $0.055 voice infrastructure, $0.015 speech, $0.04 for GPT 5.1, and Twilio’s $0.0305 to a UK mobile, about $0.14 a minute. Seven hundred and fifty minutes is roughly $105. The text fallback to the 875 calls that did not reach a person costs $0.056 each at Twilio’s UK rate, about $49. Add a number at $2.50 to $3.50. Call it $160 a month in usage, before the cost of building the agent and the fallback, and before any human time spent reading the summaries.

What that buys is about 125 conversations and, if nineteen in twenty hold as ours did, roughly 118 written, scored assessments a month, plus 875 people who received a text within a minute of the missed call. Judge the system on those two numbers and on what they led to, never on 1,000 calls made. If your volume is 100 calls a month rather than 1,000, the usage cost is trivial and the build cost dominates; if it is 10,000, concurrency charges and the model choice start to matter, and the audit should model both.

Do not use AI calling for cold marketing to individuals. Regulation 19 requires specific prior consent for automated marketing calls, the TPS removes most consumer numbers from live calling, and the ICO’s penalties are real. If the person has never asked to hear from you, a machine should not be the one to call.

Do not use it where a text would do. The Cochrane review of appointment reminders found a call and a text performed the same (risk ratio 0.99) and the text was 55 to 65 per cent cheaper per attendance. For a simple confirmation, a reminder or a yes-or-no question, a message is faster, cheaper and more polite than a call from a machine.

Do not use it for a small number of high-value conversations. The field experiment found AI interviews were rated significantly less natural even where they produced better hiring outcomes. If you make twenty calls a week and each one matters, a person should make them, with the AI writing up the notes afterwards. AI calling earns its place where the volume is high, the timing matters and a human cannot be awake for every application. If you are past those three tests and buying, ask these.

The outcome shares, the median conversation length and the share of conversations producing a usable assessment are from our own production platform over a ten week period and are published as shares only. The field experiment figures are quoted from the arXiv abstract and full text of Jabarian and Henkel (2026), a working paper not yet peer reviewed. The regulatory content was read at source: PECR regulations 19 and 21 on legislation.gov.uk, the ICO’s Guide to PECR, direct marketing guidance and penalty notice, the TPS website, and Ofcom’s persistent misuse policy. Prices were read on Retell, Vapi, ElevenLabs and Twilio’s own pages in September 2026, in US dollars as published. The worked example uses a stated hypothetical volume and no client data. First published 9 September 2026, revised 14 September 2026.