How to Use Claude Personal Finance Without Trusting the Wrong Total

Claude Personal Finance lets eligible US subscribers connect accounts through Plaid and ask about their spending. It cannot move or invest money. The useful first task is a spending audit with a total you can verify, rather than asking a chatbot how much you can afford.

Anthropic announced the beta on November 5 for Pro, Max and Ultra subscribers in the US. It can read selected accounts' balances and up to two years of transactions, and it does not see your bank password. This guide is based on the published documentation, not a test with a live bank account.

Start with a small, answerable question

I would start with one account and a completed month. That gives you a statement to compare against and avoids turning the first experiment into a full picture of your finances. Choose a question such as how much you spent on a particular service, rather than a broad request to optimise your life.

To connect, open Personal Finance in the sidebar, choose Get started and follow the Plaid flow. Review the requested access and select your accounts yourself. The rollout is gradual; Free, Team and Enterprise accounts are excluded, and only US banks are supported. Readers in India should not expect this beta to connect their Indian accounts.

The distinction between access and usefulness matters here. Connecting more accounts makes the answer look more complete, but it also increases what you must reconcile. Start with enough data to answer your question, then expand only when you have a reason.

Ask for the transactions behind the answer

Here is an original prompt you can adapt. It is a suggested workflow, not a tested guarantee of how Claude will respond:

Review spending in the selected account for October 1–31. Show the transactions behind each category total, with date, merchant and amount. Separate purchases, refunds, account transfers and card payments. List anything you cannot classify confidently. Tell me which accounts and dates the answer covers before giving a total.

This is deliberately narrower than asking where your money went. A category label can hide several decisions: whether a refund belongs to this month, whether a transfer is an expense, or whether a merchant name refers to a subscription. You want to see those decisions before relying on the result.

A total you cannot trace back to rows is unfinished work. If the explanation is unclear, ask for the disputed transactions separately and check them in your bank's own records. Do not settle a mismatch by asking the model to be more confident.

Reconcile one worked example

Consider this fictional set of transactions across a current account and a credit card. The numbers demonstrate the accounting check; they are not output from Claude or anyone's real accounts.

Transaction	Amount	Treatment in this example
Grocery purchase on card	$120	Purchase
Software purchase on card	$30	Purchase
Refund for part of the grocery purchase	$20	Reduces purchases
Payment from current account to card	$130	Pays the card balance

For a spending view that includes both accounts, net purchases are $120 + $30 − $20 = $130. Adding the $130 card payment again would report $260, even though the payment settles purchases already counted. A cash-flow view of the current account asks a different question and may correctly show that $130 leaving it.

That is why the prompt should name the question as well as the dates. “How much left my current account?” and “How much did I spend across my accounts?” can have different correct answers. Check the underlying treatment before concluding that either number is wrong.

For a real audit, compare one category first. Record the bank's total, Claude's total and the unresolved difference. If those agree, check another category. If they disagree, investigate the rows instead of asking Claude for a more elaborate dashboard.

Check data coverage before making decisions

Anthropic's setup and troubleshooting guide warns that balances can lag by a day or two, unavailable banks can disappear from spending figures, and duplicate connections can double-count accounts. Linked data is unavailable in scheduled tasks, Cowork and Claude Code.

For the audit, keep a short note beside the answer: accounts included, period covered, last bank check and unresolved transactions. If an account is missing, mark the result incomplete rather than interpreting the smaller total as an improvement. An attractive chart does not tell you whether its inputs were complete.

This is also where I would stop the experiment if it became more work than checking the statement. The point is to reduce the effort needed to understand spending. Repeatedly debugging an unreliable total defeats that purpose, however fluent the explanation sounds.

Separate privacy, access and deletion

Anthropic says conversations inside the Personal Finance hub are excluded from model training; outside that hub, your privacy settings apply. That is a specific boundary, not a promise that every conversation containing financial information has identical treatment.

Unlinking stops future bank reads but leaves existing chats and artifacts. Shared links can expose financial figures. Unlinking the last bank also resets the dashboard irreversibly. Use Manage linked accounts to unlink, then review saved and shared material separately.

The consumer retention policy says deleted conversations disappear from history immediately and are removed from backend storage within 30 days, subject to exceptions such as legal requirements and policy enforcement. Do not treat pressing delete as proof that every retained copy disappears instantly.

Before starting, decide what you will keep after the audit. My preference would be a private reconciliation note with the minimum detail needed, rather than a shareable artifact containing transaction history. Check any export before sending it to someone else, even if the intended purpose is only to discuss the layout.

The useful measure is whether you can explain the result yourself. A good spending assistant should help you reach that point with less work. If the answer only feels trustworthy because it is neatly written, keep the bank statement open.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain what their products can actually do, including the limitations users need to understand before trusting them.

Book a 30-minute call.