What a real nonprofit's ledger can't tell a keyword list

After tuning ai-bookkeeper's keywords on a makerspace's bank exports, the remaining misses are things no text rule can know.

Tools and projects · 1 min read · Source
ai-bookkeeperbookkeepingevaluation

ai-bookkeeper's built-in keywords now suggest an account for 72.7% of the 402 lines in a San Francisco makerspace's public 2019-2023 bank exports (right of those suggested: 96.0%). That set is one organisation, and the keywords were tuned on it, so read the number as a ceiling for that data, not a forecast for yours.

The last fix this pass was a bank's DEBIT-ANALYSIS CHARGE, now a bank fee.

What is left is more interesting than what was fixed. Of 76 scored lines with no suggestion, 48 are CHECK# XXXX CHECK WITHDRAWAL, which this organisation books as rent, and the rest are mostly DEPOSIT - BRANCH and ACH CREDIT <sender> lines booked as income. Nothing in those descriptions says so. A check could be rent, a refund or a vendor. That is the job of the payee memory and the review queue, and no keyword list should guess.

Data is fetched at run time and not copied (eval/queeriouslabs-bank.mjs).

More in Tools and projects

All writing · All in this topic · All tools and research