AP Auto-Recon (with a human queue)
Time saved
14h / month
Stack
QuickBooks / Bank CSV exports / Anthropic Claude / Slack escalation channel
Status
In production
01 / Problem
Bookkeeper hand-matches 200+ invoices to bank lines every month. Books close on day 9. The work is boring, error-prone, and burns out the person doing it.
02 / Trigger
Nightly job picks up new invoices + the day's bank lines and runs the match.
03 / Workflow
- Normalize vendor names (fuzzy) and date windows on both sides.
- Score every (invoice, bank line) pair on amount + date proximity + vendor similarity.
- Auto-commit matches at ≥ 92% confidence and log the reasoning.
- Suggest matches at 60–92% with a one-click approve button.
- Park the rest in a human queue with top-3 candidates pre-attached.
04 / Result
Books close by day 3. Human queue went from ~200 items to ~16 weird ones. Bookkeeper found two duplicate vendor entries hiding in the system for 18 months.
05 / Gotchas
- ! The 92% threshold isn't magic — it's tuned to "wrong less often than the bookkeeper would catch in review." Lower it and trust evaporates.
- ! Don't auto-commit anything that touches a tax-relevant ledger without an audit log the accountant can read.
- ! Vendor name normalization is where the real work is. The model is the easy part.
§ Notes from the build
Why filter, not replace
Replacing the bookkeeper was never the goal — and would have been a bad goal. The goal was to filter her queue so the work she does is the work that needs her judgment.
The system commits the boring matches. She handles the weird ones. She finds duplicates and frauds the boring matches would have buried. Net result: she’s faster, and the books are more accurate.
On the human queue
The queue isn’t a punishment for the model failing. It’s the system’s most valuable output. Each item arrives with the top three candidate matches and the reason each was rejected. She decides in seconds, and her decision feeds back into the matching rules.
That feedback loop is the difference between a system that gets quietly worse over time and one that gets quietly better.
Related field notes
All notes →№ 01
The recon that runs itself (almost)
200 invoices, one bookkeeper, nine days to close. Now the easy 92% match themselves and only the weird ones reach a human.