Reef TRHBy the first week of the month, a bookkeeper at the firm we worked with has the same problem forty times over. Forty clients' paperwork has arrived, and no two piles look alike: a supplier invoice as a phone photo, a bank statement as a PDF, three receipts forwarded in a WhatsApp message, a spreadsheet a client built themselves. Somewhere in that mess is a clean monthly report, and the old way to reach it was to type the whole thing in by hand.
"AI for accountants" gets sold as a box that does the accounting. This system is smaller and far more specific than that. It does not touch the accountant's judgement. It handles the reading, sorting and assembling that used to sit between the paperwork landing and the accountant actually looking at it. Here is that work, step by step.
It reads whatever the client actually sent
The first job is the one nobody wants: turning a photo of a crumpled invoice into usable data. A document-understanding model reads each file however it arrives, a scan, a photo, a PDF, a forwarded email, and pulls out what matters, the supplier, the date, the amount, the tax. It does not ask the client to send things in a tidy template, which is the entire point, because they never will.
It sorts each line the way the firm would
Reading is not enough; every expense still has to land in the right place. A categorisation engine, trained on the firm's own history, proposes where each line belongs in the ledger and learns from every correction an accountant makes. After a few months it stops guessing and starts matching how that particular firm books things, which is not always how a textbook would.
Nothing reaches a client on the model's say-so. Anything read with low confidence, an odd amount, a half-legible document, a category it is unsure about, is flagged and held for a person. The accountant reviews, corrects and signs. The AI assembles a finished draft; a human still owns the number that goes out. That guardrail is what makes it safe to run on real books, the same discipline behind proving an AI is right before you trust it.
It assembles the report, with the sources attached
Once the lines are read and sorted, the system builds the structured monthly report for each client and keeps every figure linked to the document it came from. So when a client asks what a charge in March was, the answer is one click away instead of an afternoon of digging. The report that used to be built from scratch at month-end now arrives mostly finished, waiting for a review rather than a rebuild.
What actually changed
The firm runs this across dozens of clients now, and the honest version of the result is quiet. The accountants did not lose their jobs; they lost the data entry. The hours that went into typing and chasing paperwork went back into the part clients actually pay for, looking at the numbers and telling them something useful. It is the clearest answer we have to the question every firm asks first, whether AI is coming for their people. The full build sits in the monthly-reporting case study.
The pile still arrives every month, as messy as ever. The difference is that by the time it reaches the accountant, it is already a report, sourced and sorted, waiting for the one thing the software was never built to do, which is decide what it all means.
Common questions
Does the AI do the accounting?
No. It reads, sorts and assembles the raw material into a draft report, and the accountant reviews, corrects and signs every one. It removes the data entry, not the judgement.
Do clients have to send documents in a set format?
No, and that is the point. It reads photos, scans, PDFs and forwarded emails as they come, which is exactly why it survives a real firm's messy intake.
What happens when it gets something wrong?
Low-confidence reads are flagged and held for a person before anything reaches the report, and every correction teaches it, so it fits your firm's way of booking over time.