The three ways it fails
Categorization. The accounting system's built-in assistant suggests an account for each transaction and is wrong most of the time, so the bookkeeper reviews every suggestion and the "automation" adds a step. Owners describe this one with real anger; they were promised less clicking and got more.
Matching. Payouts from a card processor arrive as one number; the invoices they cover are many. A model guessing which invoices a payout covers gets it right often and wrong sometimes, and "sometimes" is the whole problem at month-end.
Totals. A model asked to summarize a spreadsheet will produce confident totals that do not foot. The person who catches it never trusts the tool again, correctly.
Why a prompt can't fix it
Language models are good at language: reading a document, drafting a reply, extracting fields, classifying a description. They are not a ledger. Anything that moves money needs three things a model does not have on its own: a source of truth to reconcile against, a rule for what happens when it disagrees, and a log a human can audit. Those are engineering, and the rules are accounting.
What works instead
- The model does the reading; the code does the arithmetic. Extract the invoice number from the PDF with a model. Add the amounts with code. Never the other way round.
- Every number is reconciled to a source. The payout matches invoices whose sum equals the payout, or it does not post. The bank feed line matches a transaction, or it lands in an exceptions queue.
- Exceptions go to a human, never silently through. The automation's job is to shrink the exceptions queue every month, not to hide it.
- Parallel-run before cut-over. The system runs beside the human process for a full cycle. When they agree, and only then, it takes over.
What it looks like from the owner's chair
The bookkeeper stops clicking OK five hundred times and starts reviewing a dozen exceptions. The close goes from eight days to two. Invoices get chased on day 31 and day 45 by a system that never forgets and never sounds awkward. And the totals foot, every month, because arithmetic is not something we asked a language model to do.
If you tried AI and it didn't work, the assessment will tell you which of the three failures you hit and what the fix costs. Usually less than the last attempt.