Anthropic announcement and Help Center documentation published November 5, 2026; US beta rollout began the same week
Claude Can Read the Bank Account Now: a Read-Only Finance Beta With Its Caveats Printed in Full
Anthropic launched Personal Finance in Claude on November 5: a beta that lets subscribers link bank, credit card, loan, and investment accounts through Plaid and ask questions about balances, transactions, and spending in plain language. The boundaries are stated up front and are the substance of the launch. Access is read-only; in the Help Center's words, Personal Finance can't move, manage, or invest your money, and it is not designed to give investment advice. Sign-in happens on Plaid's screens, so Claude never asks for or sees a bank password. For each account a user chooses to share, Claude can read the bank and account names, account type, the last four digits, balances and credit limits, and up to 24 months of transactions; for investment, loan, and card accounts it reads balances and credit limits only, not holdings, trades, rates, or terms. Up to 16 US banks can be linked. The beta is limited to Pro, Max, and Ultra plans in the United States and is rolling out gradually; Free, Team, and Enterprise plans are excluded. The privacy terms are carved carefully: conversations in the Personal Finance hub are not used for model training, and chats that use linked accounts are kept out of memory, chat search, and the monthly recap. The counterweights are equally explicit in the documentation: figures stay in chat history until the chat is deleted, shared links and published artifacts can expose account figures to anyone who opens them, and unlinking a bank stops future reading but does not remove figures already present in chats. Anthropic also warns, in its own words, to check important figures against bank records: balances can be a day or two old, an account linked through two logins can be counted twice, and if Claude cannot read one linked bank, the spending totals silently omit it. Linked accounts cannot be used in scheduled tasks, Cowork tasks, or Claude Code. Anthropic says it is exploring future work alongside licensed financial professionals, which this beta is not.
Executives
For financial institutions, this is the shape of AI distribution arriving at your customer base: a general assistant reading account data through an aggregator, inside controls your customers chose, outside your own channels. The design choices here, read-only scope, no transactional authority, aggregator-mediated credentials, published data inventory, are a workable reference for what good looks like in third-party AI access, and they are worth measuring your own partners against. For everyone else, the enterprise angle is policy: this is a consumer beta excluded from Team and Enterprise plans, so personal use on personal accounts is the intended lane, and staff connecting work-related financial accounts would be outside it.
CIO / CISO
If your organization is assessing this class of feature, the documented control set is a solid starting checklist: scoped read-only access through a broker, no credential handling by the model vendor, per-account user selection, revocation in two places, and exclusion of linked data from memory, search, and automated tasks. Then test the residual risks the documentation itself names: financial figures persisting in chat history, artifacts, and shared links after access is revoked, and silent incompleteness, where a total that omits one account looks authoritative. Those two, retention after revocation and confident partial data, are the findings to carry into any comparable AI data-connection review.
Engineers
Note the integration details that carry the risk: transaction records include the bank's own description fields, and Anthropic flags that a full account number could still surface if a bank places one in an account name or transaction description, so field-level assumptions deserve verification in any similar build. Also note the consistency model: balances may lag a day or two, duplicate connections double-count, and unreadable sources drop out of totals without an error surface. If you build aggregations on aggregator data, those three behaviors, staleness, duplication, and silent omission, need explicit handling and visible flags, not footnotes.
Everyone (non-IT)
If you try the beta, three habits keep it safe. Connect only the accounts you actually want to ask about, since you choose them one by one. Never share a chat link or publish an artifact from a finance conversation, because anyone opening it can see the figures in it. And check any number that matters against your bank's own app before acting on it, exactly as Anthropic advises: the totals can be a day behind, can count an account twice, or can quietly leave one out. Unlinking is easy and stops new reading, but it does not erase numbers already sitting in old chats, so delete those chats when you are done with them.