Hot AI News

AI phone-calling agents: How Gemini and Duplex call businesses for you

An AI phone-calling agent rings a business for you, talks in real time and brings back the answer. Here is how the pipeline works, where it started and what still goes wrong.

AI phone-calling agents: How Gemini and Duplex call businesses for you
Photo illustration: Aiholics

An AI phone-calling agent is software that dials a real business, holds a spoken conversation and reports back what it learned or arranged. It does the job a person would do with a phone: ask about a price, book a slot, wait on hold. Google’s Gemini Tasks and its earlier Duplex project are the best-known examples.

One errand, start to finish

Say you need a quote for a car service from three dealerships. You type the request once. The agent finds the numbers, calls each shop and asks the same questions. It waits through hold music. Then it sends you a short comparison.

Google describes this kind of flow in the update covered in our report on Gemini Tasks going free. Per the company, Gemini can phone several businesses at once and says up front that it is an AI. Those are Google’s own descriptions, not independent test results.

The appeal is easy to see. Phone calls are slow, repetitive and often happen during working hours. A machine can make three calls in parallel without getting bored, and it does not mind a ten-minute wait. The person only sees the result.

Advertisements

Listening, deciding and speaking in a loop

A touchscreen telephone on a restaurant counter displays speech waveforms and decision panels connected in a loop.
Photo illustration: Aiholics

A call is a loop with three jobs. First, speech recognition turns the other person’s voice into text. Second, a language model reads that text, the call history and your goal, then decides what to say next. Third, a speech synthesizer reads the reply aloud.

That loop has to run fast. Humans expect an answer within a fraction of a second, and long pauses make a caller sound broken. Google’s 2018 Duplex write-up said some replies used faster approximations to answer in under 100 milliseconds.

Speed is only half the problem. The agent also has to judge turn-taking. When the other person pauses, are they finished, or just thinking? If the agent jumps in too early, it talks over them. If it waits too long, the line goes quiet and the human may hang up.

Interruptions add a second challenge. A shop employee may cut in with a question while the agent is mid-sentence. A good system stops talking, listens and adjusts its answer instead of finishing its script.

Newer systems can skip the text step and work on audio directly, an approach covered in our guide to speech-to-speech models. The goal is the same: sound natural while staying on task.

From Duplex in 2018 to Gemini Tasks today

Google announced Duplex in May 2018. Its core was a recurrent neural network, a model that reads a conversation step by step. It took speech recognition output, audio features, the history of the call and the task details, according to Google’s research blog.

Duplex worked only in narrow, closed domains such as booking an appointment or checking holiday hours. Google wrote that it “can only carry out natural conversations after being deeply trained in such domains.” Anything outside that scope was not supported.

The newer products rest on large language models, which handle open-ended wording far better than a model trained for one task. In 2025, press coverage reported that Google Search let users ask an AI to check prices and availability with local businesses in the U.S.

Gemini Tasks extends the idea. Calling sits beside shopping, bills and email in one assistant. The agent can pull context from your inbox or calendar, make the call, and then use that answer in the next step, such as a purchase.

Advertisements

Telling the other side it is a machine

Duplex drew criticism in 2018 because its voice added pauses and fillers like “uh” that made it sound human. Google’s own post said it wanted callers to be clear about the intent of the call and that the right approach was still being tested.

Today the norm is disclosure. Google says Gemini identifies itself as an AI when it calls, and that businesses can opt out of these calls. Reporting also says calling is unavailable in a handful of U.S. states.

Opt-outs matter for a practical reason. A shop that gets many automated calls may find them a nuisance, and a front desk may prefer to talk to people. Rules on automated and recorded calls also differ by country and state, so a feature that works in one market may be limited in another.

Where the approach still breaks

Phone audio is poor. Fast speech, accents and background noise cause recognition errors. Ambiguity is a classic trap: Google’s 2018 post noted that “Ok for 4” could mean a time or a party size.

Phone menus and hold queues add more risk. The agent must decide when a human has picked up and when to give up. Business policy matters too: some staff refuse to talk to a bot, and the call ends without an answer.

Duplex used human operators as a fallback. When it flagged a task it could not handle, a person took over. Google’s Tasks post admits that tasks can fail, but, as our report notes, it does not say how often.

  • Accuracy: a misheard time or price can lead to a wrong booking.
  • Consent: rules on recorded or automated calls differ by country and state.
  • Fairness: shops that ignore AI callers may lose customers who never phone themselves.

There is also a quieter risk. If an agent asks for a price and the business gives a vague answer, the summary sent back to you may sound more certain than the call was. Users should treat a quote from an agent as a lead to confirm, not as a binding offer.

Advertisements

What to watch next

Three things are worth tracking. The first is published success rates, which Google has not shared. The second is how opt-out tools evolve: businesses may get a standard way to block or route AI callers. The third is the move from calls that gather information to calls that commit money.

That last step ties into agentic payments. Google says Tasks can use a spending cap and ask for confirmation before it buys. Whether those safeguards hold up when a live call goes off script is something independent testers still have to show.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Daniel ReedAI Research, Safety & Ethics Analyst

Daniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like