What AI really does in accounting

Guide · 7 minute read

Accounting software has been advertising artificial intelligence for about a decade, and for most of that time the feature underneath was a rules engine with a good marketing budget. That changed. Reading a crumpled receipt photo, pulling a supplier name and an amount out of it, and proposing the right expense account is now something software does well enough to rely on. Knowing which parts moved and which did not is what saves you money.

The three tasks where it pays off

Reading documents

Receipt and invoice scanning is the clearest win. A model reads the image, returns supplier, date, total, VAT amount and often the line items, and the form comes back filled in. In Israel this matters more than in most places because a lot of small suppliers still hand out thermal paper receipts that fade within a year. Scanning at the moment of purchase turns a shoebox problem into a five second problem.

Accuracy is high on the fields that are printed clearly and lower on handwriting, faded ink and unusual layouts. A system that shows you what it read and lets you correct it before saving is doing this right. A system that saves silently is trading your accuracy for its demo.

Categorization that learns from your own history

The second time you buy from the same supplier, the account should already be filled in. This is not a hard machine learning problem and it does not need a large model, but it removes the most repetitive part of monthly bookkeeping. The categorizer demo on this site shows the shape of the output with a keyword table standing in for the model.

Asking questions in plain language

Reports answer the questions someone anticipated. Everything else used to mean exporting to a spreadsheet. Being able to ask how much you spent on subcontractors last quarter, or which customer is slowest to pay, and get a number back from your own books, is the feature that changes how often people actually look at their data. The requirement is that the answer comes from a real query against your records, not from a model summarizing a description of them.

Where the claims get thin

Anything described as fully automatic bookkeeping deserves a slow read of the fine print. Bank feed classification still needs review, because the same 340 shekels at the same supermarket is groceries one week and a client meeting the next, and no model can see the difference from the transaction line. Automatic VAT filing usually means the software prepares the numbers and you press submit, which is the correct division of labor and a much smaller claim than the ad implies.

Predictive cash flow is the one to be most skeptical about. With a few dozen invoices of history, a forecast is arithmetic on your existing payment terms wearing a nicer chart.

What should never be automatic

Issuing a document that carries legal weight. Once a tax invoice is issued it cannot be edited, deleted or renumbered, it becomes part of a numbered sequence with no gaps, and in Israel it may need an allocation number from the Tax Authority to be valid for the customer. That is a decision with consequences, and it belongs to a person who is looking at the screen.

The design rule worth demanding from any vendor is simple. AI produces a suggestion, a human accepts it, and only then is anything written to your books. If a feature writes a fiscal record without a person confirming it, the vendor has chosen a demo over your audit trail.

How to evaluate a tool in ten minutes

1. Photograph your worst receipt, the faded one, and see what comes back.
2. Check whether you can see and edit every field before it saves.
3. Ask the assistant a question whose answer you already know, and verify the number.
4. Try to edit an already issued invoice. If the software lets you, walk away.
5. Export your data and see what format you get.

The fifth one filters out more vendors than the other four combined.

Next: Israeli VAT without the jargon or the invoice checklist for freelancers.