An engineer in Japan who runs a few small businesses had Claude Code build a re-runnable reconciliation pipeline for his books and measured how it did on one month of 200 transactions.
- First reported
- dev.to · 25 Jun 2026
- Added
- 13 Oct 2026
- Source checked
- 13 Oct 2026
What they did
The pipeline pulls the month's transactions from his accounting platform over MCP and matches each bank row to a receipt or invoice by amount and date. For every row it proposes a category with a one-line reason, and it flags rows it is unsure about instead of guessing. He then reviewed the flagged rows and spot-checked the rest.
What happened
The poster says 189 of 200 transactions were categorized correctly. Of the 11 errors, 7 were already among the 18 rows it had flagged, and his spot-checks found the other 4.
Keep in mind
Errors came from intent it could not see, missing local tax rules and multi-purpose payments, so a human pass is still needed. The accounting platform is not named, no time saving is reported, and the post ends with a plug for his book.
Try it with your agent
Write a script I can rerun each month that pulls transactions for <month> from <your accounting system or export file>, matches each bank line to a document in <your receipts folder> by amount and date, and proposes a category from <your chart of accounts> with a one-line reason. Put every row you are not confident about in a separate review list. Do not post or change anything in the accounting system.
Replace the <angle brackets> with your own details.
Original source
dev.toUser storySomeone who uses the agent described their own setup. Figures are theirs.
Tools for this kind of work
- RepoSearch and update Jira and Confluence from your agent
- RepoQuery your business databases safely with MCP Toolbox
