deniz.in

Markets

Weather

Loading weather

· via Google AI blog

Google's Gemini can now make calls, pay bills and complete tasks for users

Google says Gemini can now phone businesses, analyse bills and complete purchases on a user's behalf, in a free update rolling out across six countries on the new Gemini 4 Hydrogen model.

Google's Gemini can now make calls, pay bills and complete tasks for users

What's rolling out

Google has started rolling out a major update to Gemini that turns the assistant into an agent capable of making phone calls, digging through bills and completing purchases on a user's behalf. According to the Google AI blog, the update is free, runs on a new model called Gemini 4 Hydrogen, and is reaching a user base Google sizes in the hundreds of millions across the United States, Japan, Korea, Canada, Australia and India. A set of more advanced capabilities debuts first in the U.S.

Calls, bills and checkouts

Google illustrates the feature set with everyday errands. In one scenario, a user asked Gemini to book a service appointment for the family minivan. Drawing on what Google calls Personal Intelligence — context such as the family's 2019 Honda Odyssey and home address — the assistant located nearby dealerships on Google Maps, rang two of them at once, waited through the hold queue, found a free slot on the user's Google Calendar and filed the booking in under five minutes. On the receiving end, Gemini identifies itself as an AI at the start of the call; businesses can opt out of such calls, and the assistant will not disclose payment or private account details unless the user says so.

Another example targeted a surprise $389 electricity bill. Granted access to the utility portal through a new Google Password Manager integration, Gemini analysed past statements in the user's email and traced the cost spike to a heat pump switching on during peak pricing hours.

For shopping, Gemini compares products via Google's Shopping Graph, asks follow-up questions about details like wattage and colour, and checks out using a new Google Pay integration — in Google's example it also surfaced a discount code worth $10.

Payments, apps and messaging

Purchases now run through the Agent Payments Protocol (AP2), with Google working with Mastercard, Visa and, described as coming soon, American Express to secure transactions with digital tokens on supported cards. Card rewards, cashback and purchase protections carry over, users can set a hard spending cap in advance, and receipts and tracking information flow into Google Wallet — which Google says is expanding to iOS as of the announcement. Google also plans to integrate the Universal Commerce Protocol (UCP) next to speed up the path from product discovery to checkout.

Beyond payments, Gemini connects to more than 60 third-party apps, including Instacart in Google's grocery example, and developers can apply to join the Connected Apps Platform. A new RCS integration lets users assign tasks over text message, including by photo: Google describes someone sending a picture of an empty cereal box with the instruction to add it to the family's grocery order.

Isolation and oversight

According to the Google AI blog, each task executes in Google Cloud's Agent Compute Engine, inside a hardware-isolated virtual machine dedicated to a single account and never shared across users. Google says these sandboxes are regularly tested by its security red teams, and that prompt-injection filters scan search results and web pages alongside Safe Browsing restrictions on which links the agent may open.

Human oversight is built in. Users can interject mid-task, Gemini is designed to pause for confirmation before clicking "Buy", it asks clarifying questions when instructions are ambiguous, and it supports "Always allow" rules for routine actions like adding an item to a grocery cart. Users can adjust task settings, disconnect apps or delete chat history, and Google is inviting researchers to probe the system through its AI Vulnerability Reward Program.

Google concedes the system still stumbles. Tasks can fail or miss the mark, and because websites were not built for AI agents, Gemini sometimes completes most of a checkout and then hands the screen back to the user for the final click. The company says it is partnering directly with major websites and security providers to help the web adapt.

Why it matters

This is one of the clearest signs yet that consumer AI agents are graduating from demos to shipped products at scale. Google is handing an agent the ability to speak on the phone, hold credentials and spend money for a potential audience of hundreds of millions, and it is leaning on its own rails — Pay, Wallet, Password Manager, the Shopping Graph — as infrastructure competitors will find hard to replicate. The caveats Google itself raises, from flaky web navigation to prompt-injection defences, are the real test: whether agentic assistants become trusted, routine tools or remain an occasional convenience will depend on how reliably and safely they perform when real money and real phone calls are involved.

  • #google
  • #gemini
  • #ai-agents
  • #payments
  • #ai-safety

Related posts