Industry News

Google Made Its Shopping AI Agent Free in Six Countries — Who Pays for Trust?

Google made Gemini Tasks free in the US, Japan, Korea, Canada, Australia and India on November 17, putting a calling, paying agent in front of hundreds of millions — while its payment rail, not its model, does the locking in.

On November 17, Google deleted a paywall that had been standing between its most ambitious consumer product and most of its users. Tasks in Gemini — the mode where you hand the assistant an errand and walk away — is now free in the United States, Japan, Korea, Canada, Australia and India. Google says the rollout reaches "hundreds of millions of people."

That sentence sounds like a pricing decision. It is really a bet about infrastructure. An assistant that can dial a dealership, read your utility statements, and check out with your card is not competing on answer quality anymore. It is competing to become the layer where your intent turns into a transaction. And once an agent holds your calendar, your passwords and your payment token, switching costs stop being about which model writes better copy.

Google published the change in a post by Josh Woodward, vice president of Labs, the Gemini app and AI Studio, under the plain title "Put Gemini to work for you." The post is unusually concrete for a launch blog: specific countries, specific features, specific failure modes. It is also, read closely, a map of what Google thinks it needs to own next.

What actually changed on November 17

Three things moved at once, and it is worth separating them.

The price went to zero. Tasks previously sat behind a Google AI Pro or Ultra subscription. Google's published list price for AI Pro has been $19.99 a month; the announcement post itself does not restate the number, but the subscription requirement it removes was the gate that kept the feature inside a paying minority.

The surface area widened. Tasks no longer only lives inside the Gemini app. In the US, users can forward email to [email protected], or text Gemini through a new RCS integration after verifying a US phone number with a supported carrier. There is now a Chat/Task toggle on the home screen: Chat answers, Task works.

The capabilities got sharper, at least in one market. Calling businesses, paying through Google Pay, and signing into accounts through Google Password Manager all start in the US. Everything underneath runs on Gemini 4 Hydrogen, the model Google introduced on November 12.

Google's own help documentation, which is more restrictive than the blog post, adds the fine print: users must be 18 or older, sign in with a personal Google account rather than a work or school one, keep Gemini history on, and can have up to 15 tasks running at once. Tasks are not available in the UK, the European Economic Area, Switzerland or Nigeria — not even on a paid plan. That is a striking gap for a product framed as a mass-market assistant, and Google has not attached a date to closing it.

Three demos, and what they give away

Woodward's post leans on his own household rather than synthetic benchmarks, which is what makes it useful. Three examples carry most of the argument.

The first is a minivan with a check-engine light. Woodward asked Gemini to call the two nearest dealerships and find an early Friday appointment within the next month. Because Personal Intelligence already knew the vehicle was a 2019 Honda Odyssey and where he lives, the agent pulled nearby shops from Google Maps, called both simultaneously, sat through the hold music, matched the offer against his calendar, and booked it in under five minutes. On the dealership's end, Gemini disclosed immediately that it was an AI, and businesses can opt out.

The second is a surprise $389 electric bill. Woodward gave the agent access to his utility portal through the new Password Manager integration and asked it to work through past statements sitting in his email. Gemini traced the spike to a heat pump switching on around 4 p.m., inside peak pricing hours. That is not a task a person performs badly out of laziness; it is a task almost nobody performs at all, because cross-referencing a portal against months of email is tedious enough that the bill just gets paid.

The third is smaller but more revealing about commercial intent. Shopping for a USB-C charger for a holiday trip, Gemini compared options through the Shopping Graph, asked follow-up questions about wattage and color, then found a discount code that knocked $10 off before completing checkout. An agent that reliably finds a working coupon is an agent that can justify a commission.

Read together, the demos describe a shift in what "assistant" means. None of the three is a question. All three are errands with a beginning, a middle, and a state change in the real world: an appointment exists, a billing anomaly is explained, a package is on its way. Google's stated goal in the post is blunt about the economics it is chasing — the company wants Gemini to "save you time and money," which is also the pitch for taking a cut of the money it saves.

The payment rail is the actual product

The most consequential paragraph in the announcement is not about models. It is about how money moves.

Purchases run through the Agent Payments Protocol, or AP2, which Google introduced in September 2025 as an open protocol for agent-led transactions. Cards are tokenized with Mastercard and Visa, with American Express listed as coming soon. Users can set a hard spending cap, and receipts and tracking land in Google Wallet, which expanded to iOS in the US the same day. Google says the next step is folding in the Universal Commerce Protocol for a faster path from discovery to checkout.

That stack — tokenized card, spending cap, signed mandate, wallet receipt — is what separates an agent that talks about buying things from an agent allowed to buy them. It is also where lock-in actually accumulates. A user can change models in an afternoon. Rebuilding a payment relationship, a password vault grant, and a calendar permission set is a different project.

This is why "free" is the least interesting word in the announcement. Google is trading subscription revenue for placement at the moment of intent. Every free user who lets the agent book, buy, or compare is a user whose transactions clear through Google's rails, and whose habit of delegating gets one repetition deeper.

The safety story, including the parts Google admits

An agent with your inbox, your passwords and your card is a liability surface, and Google's post spends real space on containment rather than burying it.

Each task runs on Google Cloud's Agent Compute Engine inside a hardware-isolated virtual machine tied to a single account and never shared with another user. Prompt-injection filters scan search results and web pages; Google Safe Browsing narrows which links the agent may open. For critical steps — sending communications, deleting data, sharing sensitive information, making purchases, submitting forms — the agent is designed to pause for confirmation or pre-approval. Routine actions can be set to "Always allow." Users can interrupt mid-task, revoke an app connection, or delete history, and Google points researchers at its AI Vulnerability Reward Program.

The honest parts matter more than the reassuring ones. Google states plainly that extensive testing has not eliminated errors: tasks can stall, or finish somewhere short of the goal. It also names a structural constraint that no amount of model quality fixes — much of the web was built to block automated traffic, so on some sites Gemini gets most of the way through a task and then hands the screen back for the final click. Google's answer is open, verifiable standards negotiated with large sites and security vendors, which is a reasonable position and also an admission that the agent era will be gated by site-by-site permission rather than solved by capability.

There is a second, less discussed constraint: calling is unavailable in five states — Indiana, Louisiana, Minnesota, Montana and Nebraska — and texting requires a US number on an eligible carrier. Advanced features are, for now, a US product wearing a six-country launch.

Why rivals should be nervous anyway

The obvious comparison is feature parity. Claude Duet, OpenAI's Dots and Meta's Muse all push in the same direction: an agent that acts rather than answers. On raw capability, the gap between these systems is measured in weeks, not generations.

Distribution is a different contest. Giving away an agent that spends money puts it in front of a free user base measured in hundreds of millions, in the six markets Google chose — four English-speaking, plus Japan, Korea and India, which happen to be among the most valuable e-commerce and messaging markets on earth. Rivals that keep their agents behind waitlists or premium tiers are running a smaller experiment with richer users; Google is running a very large experiment with ordinary ones.

The counterweight is institutional resistance, and it is already visible. Retailers and airlines have spent the past year deciding whether an AI agent is a customer or a crawler. Amazon has moved against third-party shopping agents, and United Airlines restricted AI-driven bookings. Those decisions are not about model quality. They are about who captures the margin when a purchase is negotiated by software, and every merchant that blocks an agent pushes the cost of that friction back onto the user.

That friction is the real test of the free rollout. If Tasks works on the long tail of ordinary sites — the local utility, the dealership, the small retailer — the agent becomes a habit. If it repeatedly reaches the last step and hands the screen back, users will treat it as a faster way to start an errand rather than a way to finish one.

How to judge an agent like this

Vendor demos are curated; your bills are not. The useful question is not which assistant tops a leaderboard, but what a completed outcome costs you in money, time and attention. A $10 coupon found automatically is worth more than a benchmark point. A five-minute booking that would have cost you twenty minutes of hold music is worth more than a benchmark point too.

That is the framing we use at aibookmarks.net: compare tools by cost per finished task, not by rank on a scoreboard. An agent that saves you $10 and forty minutes a week is cheap at any subscription price; one that gets 90 percent of the way and stalls is expensive even when it is free. Google's giveaway is worth testing on exactly that basis — pick three errands you actually dread, hand them over, and count how many come back done.

The bottom line

Google did two things at once on November 17. It made an agentic assistant free for a very large audience, and it put its payment, identity and commerce protocols underneath that audience before the habit formed. The model is the visible part; the rails are the durable part.

For users, the offer is genuinely useful and genuinely risky in the same breath, which is why Google's own documentation spends more words on confirmations and spending caps than on capability. For competitors, the challenge is no longer building a comparable agent. It is persuading merchants, carriers and regulators that their rails deserve the same trust.

The next twelve months will decide which of those two problems turns out to be harder.

---

Written by the AI Bookmarks editorial desk, drawing on Google's November 17 announcement and its Tasks help documentation, with context from press reporting on the Spark-to-Tasks transition and on merchant restrictions against shopping agents.

Published by AI Bookmarks · More AI news →

← Back to News