Google's Duplex Now Calls Businesses for You Inside Gemini Tasks
Google announced it on 20 November 2026, and the timing is worth noting: Duplex, the AI phone-calling system Google first demoed back in 2018, is now built into Tasks in the Gemini app. You describe a chore in plain text, and Gemini can pick up the phone and call the business for you when there is no booking form or API to use instead.
I have not tested this hands-on — it is a launch-week announcement, and the rollout details are thin — so this is a launch-day piece built from Google's own research blog and independent coverage, not a review. But the design decision at the centre of it is worth talking about, because it is the decision every agent builder eventually faces: what does your agent do when the structured channel does not exist?
The phone call is the fallback API
Here is the part that matters. Tasks in Gemini is a personal AI collaborator for real-world chores. It can analyse a utility bill, complete a transaction through the Agent Payments Protocol, and coordinate with you over email and RCS. And now it can call a business.
Read that list again and notice the pattern. Every one of those is a task where the "proper" digital channel either does not exist or is not worth building. Most small businesses will never have an API. Many do not even have a working booking form. The phone is the interface they already answer. Google's move is to treat live telephony as just another modality the agent picks when APIs and web forms fail — not as a party trick, but as the fallback that makes the rest of the agent actually useful.
That is a genuinely developer-relevant idea. If you are building agents, the lesson is not "add voice calling." It is that your agent's coverage is defined by its worst channel, not its best one. An agent that can only work where a clean API exists stops being useful exactly where real life starts.
What Google says it can do
The announcement, on Google's research blog, describes live calls powered by Gemini 4 Hydrogen and Gemini 4 Neon, unifying speech understanding, planning and dialogue generation in one stack. The history behind it is long:
- The original Duplex (2018) was constrained to narrow domains, with custom recurrent neural network models trained on anonymised, task-specific phone conversations. No pre-trained world knowledge; every task type needed its own training data.
- The hybrid era (2022 onwards) brought open-domain foundation models into live calls — LaMDA in 2022 — paired with deterministic state tracking and strict policy guardrails, because early generative models were less predictable and slower than today's.
- The current stack drops fixed verticals entirely. Tasks are described in free text, and the system generalises to new task types and new business categories without building a separate dialogue model for each one. A request like "book the window seat and ask if they have space for a stroller" does not need a restaurant-booking model; it needs a model that understands the request.
During the COVID-19 pandemic, Duplex doubled down on verifying merchant information automatically, and Google says business information sourced from those calls had been shown in Search and Maps over one trillion times, as shared in 2023. The same models later powered Pixel Call Assist features like Call Screen and Hold for Me across five consecutive Pixel generations, and in 2025 and 2026, broader agentic calling arrived in Google Search and Ask Maps, checking quotes, product availability and service openings from pet grooming to auto repair.
The technical hurdles Google names are the honest ones, and they are worth reading as a checklist if you ever build anything near telephony:
- Pragmatic grounding. The web does not teach you how a call with a busy merchant actually unfolds — how to read a polite indirect refusal, how to know what makes sense to ask at a given moment, or how to keep the call brief. Nearly a decade of goal-directed calls is the training data.
- Sub-second latency on noisy lines. Calls run over compressed phone lines against kitchen clatter, espresso machines and street noise. The stack is tuned for low-latency streaming inference, with custom acoustic and conversation-cadence models handling background noise and pauses.
The part Google got right: the business on the other end
Duplex has to work for the person answering the phone, not just the person who sent the task. Google's design choices here are the ones I would steal:
- Clear disclosure. Every call opens by clearly disclosing that it is an automated system calling from Google. Not buried in the middle of the call, not implied. Up front.
- Businesses stay in control. Merchants can opt out of automated calls in their Google Business Profile settings, or simply say "please stop calling my business" during a call.
- No setup, zero adaptation. Duplex works through the phone a business already answers. No software to install, no booking system to maintain, no staff to train. Whoever picks up handles the call the way they would with any customer, even mid-rush with a line at the counter.
- One conversation, many insights. Details gathered on one call can answer many later questions, so a business is not asked the same thing over and over. Google Maps can show details Duplex gathers, like happy hour specials, patio seating, pet-friendliness or a parking tip.
That last point cuts both ways, and I will come back to it.
What Google did not publish
Here is the honest part, stated plainly. Google's announcement says nothing about:
- No API keys, no public Tasks telephony SDK. There is no way for a developer to build on this calling stack today. It is a Google product feature, not a platform.
- No per-minute pricing, and no pricing at all. Calls are presumably bundled into the Gemini experience, but Google has not said what the limits are, or what happens when you hit them.
- No day-one country list. The announcement does not say where Tasks calling works on day one, or when it reaches anywhere else.
If you are a developer reading this as a platform announcement, it is not one. Do not plan a roadmap around an SDK that does not exist yet.
The context: Tasks went free three days earlier
This did not come out of nowhere. Tasks launched to select testers in May as Gemini Spark, a 24/7 agentic helper, became Task Mode, and was expanded to US subscribers of the Google AI Ultra tier. On 17 November 2026, Google expanded it to free users on Gemini 4 Hydrogen — TechCrunch reported it reaching hundreds of millions of people. Josh Woodward, VP of Gemini, said he used it to call local car dealerships to schedule a service appointment for his family's minivan, and to analyse his electrical bill.
So the sequence is deliberate: make Tasks free on 17 November, then announce the calling capability inside it on 20 November. Google is pushing Gemini from "the one that answers" to "the one that does things," and the phone call is the most visible proof of that shift.
The question I would actually ask
There is one thing in the announcement I keep coming back to. "One conversation, many insights" — details gathered on one call can answer many later questions. That is efficient, and it is also a data-collection pipeline dressed as convenience. A restaurant tells Duplex about its patio seating once, and that detail shows up in Maps for everyone. The business never explicitly agreed to publish it; it just answered a phone call.
I am not saying that is wrong. I am saying it is the part of this launch that deserves the most scrutiny, from developers and from the businesses answering these calls. Disclosure at the start of the call is good. But disclosure that the call is automated is not the same as disclosure that the answers may be republished. Watch how Google handles that line as call volume grows, and watch whether merchants start noticing their offhand answers showing up in listings.
For now, the practical takeaway is simple. If you have access to Tasks in Gemini, try it on something low-stakes first — a hair appointment, a restaurant booking, a quote from a local garage. Listen to how it handles the moment the receptionist goes off-script, because that is where every phone agent lives or dies. And if you are building agents of your own, ask yourself what your agent does when the API does not exist. If the answer is "it gives up," the phone call is your homework.
Would you let an AI call a business on your behalf — and if you have tried one of these calling agents, what did it get wrong that a human would not have?
Sources: Google Research blog, 20 November 2026 — "How Google Duplex evolved to also call businesses for you in Gemini" (live calls powered by Gemini 4 Hydrogen and Gemini 4 Neon; disclosure, opt-out and no-setup design; Tasks capabilities including utility-bill analysis, Agent Payments Protocol transactions, email and RCS coordination); TechCrunch, 17 November 2026 — Google expanded Tasks access to free users on Gemini 4 Hydrogen (Tasks launched May 2026 as Gemini Spark, became Task Mode, was Ultra-only). No hands-on testing was performed for this article. The "what Google did not publish" section reflects the absence of those details from the announcement, not a claim that they do not exist. Originally published on ToolsDig.

