Claude Code pulled a month of transactions over MCP and matched bank rows to receipts, proposing categories; it reconciled 200 rows but miscategorized 11, caught in a manual audit.
Claude Code pulled a month of transactions over MCP and matched bank rows to receipts, proposing categories; it reconciled 200 rows but miscategorized 11, caught in a manual audit.
Rather than pasting a CSV into chat, the author had Claude Code build a re-runnable pipeline. It pulled the month's transactions from the accounting platform over MCP, matched bank rows to receipts by amount and date, drafted a category with a reason for each row, and flagged rows it was unsure about. The author reviewed the flagged rows and sampled the rest.
It reconciled 200 transactions and miscategorized 11 (94.5%). The errors were judgment calls: 5 on business intent, 4 on local tax rules, 2 on edge cases. Seven of the 11 were already in its flagged set of 18.
Telling the agent never to guess silently turned a 200-row audit into an 18-row review plus a sample. Even so, four errors surfaced only in the spot-check.
Reconcile my business transactions for [month] using [bank export/accounting tool] and the receipts in [folder]. Match each bank row to a receipt, propose a category from [my chart of accounts], and mark low-confidence matches separately. Don't post or change anything in my books. You're done when I have a table of matches and a list of items needing my review, so I can spot-check the categories.
An app built with Claude Code extracts vendor, date, total, tax and line items from receipts, including foreign-currency ones at historical exchange rates, for freelancers and small businesses.
A tenant used ChatGPT and Gemini to review a landlord's apartment deposit dispute, identify legal violations, and avoid losing $4,200.
The poster used an AI assistant to summarize a denial PDF to draft an appeal, but it invented an appeal deadline.