anyone using AI to classify bank transactions? accuracy + workflow question
we brought on 3 new bookkeeping clients this quarter and the transaction volume is brutal. been experimenting with feeding bank CSV exports into Claude and asking it to classify each line into the chart of accounts.
results so far:
- straightforward stuff (rent, payroll, utilities) → ~97% accurate
- merchant names it can google → ~90% (it figures out "SQ *JOESCOFFEE" is meals/entertainment)
- ambiguous transfers between accounts → coin flip
- owner draws vs expenses → genuinely dangerous if you don't catch it
my current workflow: AI classifies everything, i review the flagged items + spot-check 20% of the "confident" ones. cuts classification time from ~4 hours to ~45 minutes per client per month.
question for the group: anyone built a more systematic QA layer? like a second pass where the agent checks its own work against rules you define? or am i overthinking this and manual spot-check is fine?
4 replies
i built a two-pass system in Cursor. first pass: classify. second pass: i give it a list of rules ("any transaction over $5k must be flagged", "owner's name in payee = draw not expense", etc) and ask it to audit its own output.
catches maybe 3-4 misclassifications per 500 transactions that the first pass missed. not perfect but better than one pass alone.
been doing something similar for Maltese clients. the trick i found: give the agent the PRIOR MONTH's classified transactions as context. accuracy on recurring vendors goes from ~90% to ~99% because it pattern-matches against what you already approved.
still review everything manually for new vendors. but for the 80% of transactions that are repeats? massive time saver.
my concern with this whole approach is the client data angle. are you uploading raw bank feeds with client names, account numbers etc to a cloud model? because for UK GDPR purposes that's a data processing activity that needs to be in your engagement letter.
i only do this with local-runtime models or after stripping PII from the CSV first. slower but defensible.
rachel raises a good point. but practically — the time savings are too big to ignore. we went from dreading bookkeeping clients to actively onboarding them.
my accuracy numbers are similar to sarah's. the key insight: AI classification is a FIRST DRAFT, not a final answer. treat it like a junior did the work. you still review, but you're reviewing not starting from scratch.
Sign in as a verified accountant to reply.
Sign in