Claude can now read your bank accounts. It still cannot move the money

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 5 November Anthropic began rolling out Personal Finance in Claude in beta, for its Pro, Max and Ultra plans in the US. Customers connect bank and card accounts by signing in to their bank through Plaid, so Claude never sees the password. For each connected account Claude sees the bank and account names, the account type and last four digits, balances and credit limits, and up to two years of transactions. Access is read-only: Anthropic says the feature "can't move, manage, or invest your money".

Anthropic says chats in the Personal Finance hub are not used to train its models, and that the feature has no ads or sponsored recommendations and does not offer its own cards, loans or savings accounts. It also says Claude "can make mistakes" and tells customers to check important figures with their bank. Its examples include finding charges that hit twice, unexpected fees and a bill that has crept up over the year. And it closes with an invitation: Anthropic wants to hear from "banks, wealth managers, and financial services companies" that want to help shape what comes next.

Plaid says Anthropic joins more than 800 AI companies building on its network. It says nearly two-thirds of US consumers who connect bank accounts to AI assistants through Plaid had never used a Plaid-powered personal finance app before, and that its research found 69% of US consumers willing to share personal information with AI in financial services when they see a clear benefit.

Why it matters

Read-only is a permission boundary, and it is the right place to start. It also tells you where the harder argument will come. Issue 03 reported that the biggest US banks were taking the keys back from AI agents. Read access through an aggregator, approved by the customer, is the version banks can live with. Write access, where an agent moves money, is the version that raises the questions this issue keeps returning to: who authorises, through which channel, and who pays when it goes wrong.

The first thing a customer's AI will audit is not the customer's spending. It is the bank's fees.

The near-term effect on banks is more practical. A customer who asks "anything look off?" will get a list: duplicate charges, fees and interest costs. Some of those items will be correct, and some will be the model's mistakes. Either way they will arrive at the contact centre and the disputes team. Banks should expect customers to arrive with an AI-written list and be ready to answer it item by item. That means fee descriptions on statements a model can read correctly, and staff who can explain a charge quickly. Anthropic's invitation to regulated firms is also worth taking seriously. The firms that help define how a licensed adviser fits into these conversations will influence the rules for write access later.

What I'd watch

• Whether Anthropic adds any write actions, such as payments or transfers, and through which authorisation channel.
• Whether the feature expands beyond the US, and which data-access rules it meets in each market.
• Which banks or wealth managers take up Anthropic's invitation, and on what terms.