Telemarketer / telesales agent — tasks, one by one

Building and cleaning call lists

Matching, de-duplicating and checking numbers against do-not-call lists is rule-following on structured data, and dialing software and data vendors were doing it before language models. What is left for a person is the exception: a number that looks wrong, or a contact who says they never agreed to be called.

Automating the list does not make it lawful. Who consented, and to what, is a record someone answers for, and a list bought from another organisation can carry consent that does not cover this call. This judgement describes what the tools can do, not how many small teams still keep lists by hand. A rule requiring a named person to verify consent before each campaign would bring part of this task back to people.

Delivering the pitch

The pitch is a script, so its content is the easiest part to generate, and speech synthesis can now say it in a natural voice and handle the questions the script anticipates. Two things keep a person on most cold calls. Rules on unsolicited calls in several large markets treat an artificial voice like a recorded message, which needs the listener's prior consent. And listeners hang up or complain faster once they suspect a machine. So the change has come around the person: dialers connect them only to answered calls, and on-screen prompts suggest the next line and the answer to a question as the call goes on.

The protection here is a rule and a customer reaction, not a limit of the technology. If the rules change, or in a market where they are not enforced, this task would quickly turn to automating. Augmenting is not neutral for headcount either: when a dialer and live prompts let one person get through far more answered calls, the same list needs fewer callers. A dated, large-scale deployment of voice agents making cold sales calls in a market with consent rules, with no enforcement following, would move this judgement.

Handling objections and reading the person

A cold call is a stranger asking for money or time, and the person deciding is weighing whether the caller is real and will answer for what they promise. A language model can produce a plausible reply to a stated objection; the hard part is the unstated one, and when a listener asks whether they are talking to a machine the call usually ends. Mis-selling is also a liability: a promise the product does not keep comes back as a complaint or a cancelled sale, and employers are slow to let a system make such promises unsupervised.

This protects a skill at the top of the job, not the number of people doing it: if machines and dialers take the easy calls, fewer callers are needed for the hard ones. It also rests partly on listeners distrusting machine voices, which may fade as people get used to them. Evidence that machine-made calls persuade reluctant strangers as well as people do, measured over time and counting cancellations, would move this.

Taking orders and logging every call

Card details have long been taken by keypad systems rather than read aloud, partly because card-security rules discourage agents from hearing them. Speech recognition can now transcribe the call, and a language model can fill the order form and write the call note from the transcript, which is the kind of structured extraction these systems do reliably. The person checks and corrects instead of typing.

Automatic notes remove typing, not responsibility: a wrong order or a misrecorded refusal is still the company's problem, so checking stays. This judgement does not say whether the time saved becomes fewer callers or more calls per shift; that is the employer's choice. A pattern of automatic notes misrecording consent or refusals, leading employers to give the task back to people, would move it.

Booking appointments for salespeople

Once someone has asked to be contacted, booking a time is a short, structured exchange: offer slots, check the calendar, confirm, remind. Booking links, text-messaging bots and voice agents can do this alone, and consent is less of a barrier because the person asked to hear back. Cold appointment setting, with someone who did not ask, falls back on the pitch and the objections above.

Booked is not attended: if automated booking fills calendars with people who do not turn up, employers may bring a person back to qualify leads before booking. This judgement covers the warm end of the task; the cold end is judged under the pitch. Evidence that automated booking produces clearly worse attendance than human setters would move it.

Running surveys by phone

Surveys were automated well before language models: keypad surveys after a call, recorded-voice polls, and a shift of much research to online panels. A fixed questionnaire is the kind of script a voice agent follows well, recording answers as it goes and asking follow-up questions only where the questionnaire allows.

A survey is only as good as who answers, and people who stay on the line with a machine may differ from those who stay on with a person, so research buyers who care about a representative sample may keep interviewers for some studies. Evidence that machine-run phone surveys give different results from interviewer-run ones on the same questions would move part of this back to people.

Taking over calls a voice agent hands on

Where a system opens the call, someone has to take over when the conversation leaves the script, the listener asks for a person, or the sale needs someone who can be held to it. That person also has to catch anything the machine promised that the company cannot honour.

A new task is not new headcount: one person can stand behind many automated lines, which is the point of running them. A dated deployment that shows how many people such a setup keeps per automated line would show whether this work grows or stays small.