On 24 September Google started rolling out Call for Me, an experimental Gemini feature that phones businesses on your behalf. It dials from your own number, works through the automated menu, waits on hold for as long as it takes, and then talks to whoever picks up — checking stock, asking for a quote, booking or moving an appointment. You watch a live transcript on your phone and can take over at any point.
For now it's a narrow preview: Pixel 11 owners in the US, on a paid Gemini plan, enrolled in the Phone by Google beta, and even then rolling out gradually. But the direction is not subtle. The person on the other end of a business's phone line is increasingly going to be software.
I build the other end of that call. My AI receptionists answer real business phones, and I recently built an outbound agent that places calls in twenty countries. So I've spent a lot of time thinking about what happens when both ends of the line are machines — and the answer is that most business phone setups were designed around assumptions that stop holding.
This time it announces itself
Google has been here before. The original Duplex demo in 2018 booked a hair appointment so convincingly that the conversation afterwards was mostly about the fact that the person on the phone had no idea they were talking to a machine.
Call for Me opens differently: it introduces itself as an AI assistant calling on behalf of a user, on a recorded line, before the conversation starts. That's the right call, and it's also useful to you — a clear, consistent disclosure at the top of a call is a signal your own system can listen for.
Identity is not authority
The first thing to get straight: an AI calling on behalf of someone, from their number, has their identity but not their authority.
Caller ID was never authentication; it only ever felt like it because the person holding the phone was usually the person who owned it. An assistant placing calls from your customer's number breaks that last bit of comfort. The call is genuinely from their phone, genuinely on their behalf, and there is genuinely no human confirming the specific thing being asked for right now.
So split what your phone line does into two kinds of work:
- Information and bookings — opening hours, stock, a quote, an appointment slot. Let any caller do these, human or not. That's the whole point.
- Anything with consequences — payments, refunds, account changes, personal data, cancellations with a fee. Don't complete these on the strength of the caller. Send a confirmation link or one-time code to the account holder through a channel you already trust, and let them approve it.
If your current process is "the caller knew the customer's name and date of birth, so we changed the address", that process was already weak. With assistants that hold their user's personal details, it's now an open door.
Design for a patient, literal caller
A human caller is impatient and good at guessing. An AI caller is the opposite: infinitely patient and fairly literal. That shifts several things.
Hold queues get longer. Waiting on hold costs the assistant's user nothing, so people who would have hung up after ten minutes won't. If your staffing assumes abandonment, it's about to be wrong. Offer callbacks.
Menus need to be speakable. Accept speech, not just keypad presses, and say the options plainly. "Please listen carefully as our options have changed" followed by nine choices is a bad experience for people and a slow one for machines.
Read details back, then write them down. Two systems talking can each mishear the other and both carry on confidently. Read back names, dates, times and numbers in a fixed format — "Thursday the second of October, at 3pm, for Sarah Khan" — and send the same thing by SMS or email at the end of the call. A written confirmation settles any dispute about what was agreed, whoever made the call.
None of this is exotic. Every one of these changes also helps human callers, which is the best argument for making them now.
When two voice agents meet
This is the part I find most interesting, because it breaks a quiet assumption inside every voice agent.
A voice agent decides you've finished speaking by waiting for a short silence. It's called endpointing, and it's the biggest single delay in most pipelines — I built a latency budget calculator to show how much of a response time it eats. Humans cover that pause without noticing. Two agents facing each other don't. Each one waits for silence, then takes about a second to respond, and the conversation develops a slow, stilted rhythm — or, worse, both start talking into the same gap and interrupt each other.
What helps, in practice:
- Keep endpointing tight. Every half-second of padding you added "to be safe" is paid twice per exchange when the other side is also an agent.
- Be brief. Pleasantries and filler that sound warm to a person are dead time to an assistant that only wants a slot and a price.
- Notice who you're talking to. If a caller opens by announcing itself as an AI assistant, your agent can switch to a more structured mode: skip the small talk, answer the question directly, and read back every detail. I'd build this as a small, explicit behaviour change rather than anything clever — the disclosure is right there at the start of the call.
Two disclosures, one call
If your line is answered by an AI, it should say so as well — and in several markets it has to. That means a call between an assistant and an AI receptionist opens with two disclosures and two recording notices. That's fine. It's also a reason to keep both short: nobody, human or machine, needs thirty seconds of preamble.
It's worth checking how your system handles recording consent in both directions. The assistant tells you the call is recorded on its side; you're probably recording on yours. Know what you're storing, for how long, and who can read it.
The upside is real
It's easy to read all this as a list of new problems. It's also a shift in who gets the booking.
When an assistant is phoning around on someone's behalf, the business that answers first, answers at 11pm, and confirms cleanly wins. The one that sends it to voicemail, or strands it in a menu tree, simply doesn't get the appointment — and the customer never even knows they were turned away. Businesses that already answer every call with an AI receptionist are, quite accidentally, the best prepared for callers that are AIs too.
Call for Me is a preview on one phone in one country, so there's no need to rebuild anything this week. But the changes worth making — separating information from authority, confirming in writing, tightening turn-taking, speakable menus — are cheap, and they make your phone line better for the humans still calling it. Make them before the machines are a meaningful share of your callers, not after.
References: Gemini can call businesses and wait on hold — HotHardware · Google tests Gemini calling feature for Pixel 11 — TechRepublic · Gemini calls businesses for Pixel 11 users — The Mac Observer · Call for Me US preview — Tech Journal