Claude Can Read Your Bank Account. Read the Fine Print

Anthropic launched Personal Finance in Claude today. You link your bank through Plaid and ask where your money went. You cannot use it from Canada, and if you are on a business plan you cannot use it at all. The document worth reading is the one listing what it gets wrong.

What was launched

You connect bank, card, loan and investment accounts, and then ask questions in ordinary language. Where did my money go this month. Is that more than usual. Am I spending more than I earn. Claude adds up spending across the connected accounts, sorts it by category, and will build a chart or a calendar of what is due.

Access is read-only. Anthropic is unambiguous that the feature cannot move, manage or invest money, and says it is not designed to give investment advice. Anthropic also notes there are no ads or sponsored recommendations and that it does not sell its own cards, loans or savings products, which is a real distinction from most of the category.

The availability constraints are the first thing a Canadian business reader should register.

Not on Team or Enterprise is the line that settles the business question. Those are the plans companies buy, so this is a consumer feature, and treating it as a route to automated bookkeeping would be a misreading. For that work the ground has not moved, and the honest picture is still the one in what AI bookkeeping automates and what it cannot.

The caveats are the story

Anthropic published a list of the ways the feature produces wrong numbers. Vendors do not usually do this, and the list is more instructive than the launch copy.

Balances can be a day or two old. An account you have linked through two different logins may be counted twice. And the one that deserves attention: when Claude cannot read one of your banks, the spending figures simply leave that bank out, so a budget can show more money left than you have.

That third failure is the shape worth learning, because it recurs everywhere AI reads from a set of sources. The output is not visibly broken. It is complete-looking, specific, and wrong in the direction that feels like good news. Nothing in the answer tells you a source dropped out.

Anthropic's own mitigation is the right one and it is worth copying as a habit: ask which transactions are behind a number, and check important figures against the bank's own app. Being able to ask what a figure was built from is the practical version of the discipline in knowing how clean the underlying data is.

The data scope, precisely

For the accounts you connect, Claude reads the bank name, each account name, type and last four digits, balances and credit limits, and up to 24 months of transactions with the amount, date, merchant, category and the bank's own description.

Investments, loans and cards are narrower than people assume. Balances and credit limits, rather than holdings, trades, rates or terms.

One detail is easy to miss and genuinely useful. Plaid passes only the last few digits of an account number, but Anthropic notes that if your bank has written a full account number into an account name or a transaction description, Claude may see it there. The permission model is tight and the free-text fields leak around it, which is how most data boundaries actually fail.

What training protection you get

Conversations in the Personal Finance hub are not used to train Anthropic models. Anything you share with Claude outside the hub falls under your normal privacy settings.

The protection attaches to a surface rather than to the subject matter, which is the part to internalise. Pasting a bank statement into an ordinary chat window is not covered by it. If your team is going to handle financial information in an AI tool, the rule has to name the place as well as the data, and that belongs in writing, which is the gap most businesses still have per the research on how few have an AI policy.

Read the terms yourself rather than trusting a summary, including this one. What a vendor retains, for how long, and who can see it varies by plan and by surface, as set out in what AI terms of service actually say about your data.

What this signals

Read-only aggregation through a dedicated bank connector, with a narrow declared data scope and a published list of failure modes, is a reasonable template. It is also a one-way door being opened slowly: the announcement calls Personal Finance a first step.

The question to settle before business versions arrive is not whether the connection is secure. It is what you will do differently when an AI tool hands you a confident figure assembled from several systems. Ask what it was built from, and keep a way to check it, which is the same discipline as asking what evidence supports a vendor claim rather than accepting the label.

Frequently Asked Questions