Gemini Tasks Is Free Now, and Your Bank Says You're Liable
Google's agent can now buy things with Google Pay at no cost to you. The same week, Citi told customers they own the risk if it gets hacked.

Google made its Gemini Tasks agent free this week, including purchases through Google Pay and sign-ins to saved accounts.
The same week, the institutions holding the money moved the opposite way. Citibank changed its terms to tell customers that if a third-party AI tool leaks their credentials or is hacked, they are responsible for what happens next. Both things are now true at once, and that gap is the story.
What Tasks does now
Tasks is the renamed Gemini Spark, the 24/7 agentic helper Google launched to select testers in May. A toggle on the Gemini homepage switches from "Chat" to "Tasks," and the agent moves from answering questions to doing errands: meal plans, calendar management, digging through email for recurring expenses you may not need.
US users get the consequential additions. The agent can call businesses to compare product pricing and check inventory, book appointments, purchase items with Google Pay, and sign into accounts that support Google Password Manager. You can also hand it work by emailing gemini@google.com or texting it.
Josh Woodword, VP of Gemini, said he recently used the feature to call local car dealerships to schedule a service appointment for his family's minivan, and to analyze his electrical bill. Those are the company's own examples, and they are a fair picture of the intended scope: low-stakes phone calls and document reading, not portfolio management.
Who gets it, and what changed about the price
Spark was previously limited to Google AI Pro and Ultra subscribers. The rebrand drops that gate: Tasks is now available free in the US, Canada, Japan, Korea, Australia and India, running on Google's Gemini 4 Hydrogen model. TechCrunch described the reach as free availability to hundreds of millions of people on that model.
The practical read: the agent that can spend your money is no longer behind a subscription that implied a committed user. It is a toggle, default-adjacent, on an app most Android users already have.
The banks moved the other way in the same week
Citibank, the third largest commercial bank in the United States, changed its terms of service this week to tell customers who use third-party AI tools for personal finance that anything that goes wrong is their problem. In the company's words to TechCrunch:
"As third-party AI tools become more widely available, we want customers to understand the risks associated with sharing banking credentials or authorizing those tools to access their accounts... customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents."
This is not one bank hedging. JPMorgan and Bank of America, the two largest in the country, announced similar policies last month. PNC, the eighth largest, changed its terms the same day as the Citi report. Citizens emailed customers this week with the same warning.
Citi's version goes furthest: it tells customers that if the AI tool leaks their credentials or is hacked, the customer is responsible for what hackers then do with the account. Note also what Citi carved out — agents the bank itself provides can be used without that additional liability. The liability line is being drawn around third-party agents specifically.
Anthropic's positioning on the same problem is worth reading next to this. The personal finance tool it launched earlier this month is explicitly billed as "read-only" and, per the company, "can't move, manage, or invest your money." One vendor is shipping write access to payments; another is advertising the absence of it.
The permission model, read closely
Google is not careless here, and the safeguards are specific. Google says Gemini performs tasks in a "hardware-isolated virtual machine that's dedicated to your account and never shared with other users." You can give it instructions mid-task. And for sensitive actions such as purchasing an item, Google says it stops and asks permission before following through.
That is a reasonable design. It is also a design whose weakest point is the human approving the prompt — the same failure mode the banks are now pricing into their terms. A confirmation step protects you from an agent acting alone. It does nothing about an agent that has been talked into asking you for the wrong thing, and nothing about credentials held by a third party that is later breached.
The retention change in the background
In the same news cycle, Google extended how long it keeps temporary Gemini chats and conversations made with history turned off — to around two months, up from 72 hours, rolling out globally over the coming weeks. The previous 72-hour figure came from Google's Gemini Apps Privacy Hub as of an archived version from November 12.
For comparison, OpenAI says it may keep temporary ChatGPT chats for up to 30 days, and Anthropic retains incognito Claude chats for 30 days by default. Google's new window is roughly double either. Google did not answer questions about what prompted the change, pointing instead to a blog post citing responding to users, processing feedback, and protecting its services; its Privacy Hub also cites potential system failures, and says temporary conversations are not used to train models.
One genuine improvement arrived with it: chat history and AI training are now separate settings, where both were previously tied to a single "Keep Activity" control. Users who had Keep Activity off get both new settings off by default. If you had it on, your history stays on and you choose training separately.
Retention matters more once the agent is transactional. A chat log is a record of what you asked. A task log is a record of what was bought, which accounts were opened, and which businesses were called on your behalf — held for two months, including in the mode users pick when they want less kept.
Should you turn it on?
- For calls, appointments and reading your own documents: this is the use case Google demonstrated, and it carries no credential risk beyond what the app already has. Low downside.
- For anything touching a bank or brokerage: read your own institution's terms first. Four of the largest US banks have now moved liability to the customer for third-party agents, and Citi's wording covers breaches of the tool, not just its mistakes.
- For account sign-ins via Password Manager: treat each one as granting standing access, not a one-time login. That is the credential surface the banks' terms are written about.
- If you use temporary chats for privacy: recheck your settings after the rollout. The retention window changed and the two settings split; defaults were re-derived from your old Keep Activity state.
What to watch
The open question is not whether agents can transact — Tasks demonstrably can. It is who absorbs the loss when one goes wrong, and right now the answer from the four largest US banks is: not them. Watch whether a major platform offers an indemnity for agent-initiated transactions, whether regulators treat these terms changes as a consumer-protection matter, and whether other banks follow PNC and Citizens in the next few weeks. Until one of those moves, free agentic purchasing and customer-side liability are shipping together.
More from DangMua