AI bookkeeping for Saudi companies, with a licensed accountant on the end of it

The daily loop

Every working day, documents are pulled from your accounting system and your payment providers. They are mirrored into a separate double-entry ledger, so there is always an independent set of books to compare against the one you keep. Matching runs first: reference against reference, amount against date, and payout totals broken back into their orders and fees.

What matches becomes a proposed entry. What does not becomes one short question, sent on WhatsApp, written for a business owner rather than for an accountant. A transfer arrived, no invoice matches it, was it the February retainer. You answer, the entry is draft-posted, and your accountant approves it with the rest.

Where the model earns its place

Reading documents is the honest use. A supplier sends a PDF with the VAT number in a different place every month, an Arabic delivery note arrives as a photograph, an expense claim has the total handwritten. A model reads those far better than a rule ever will, and the extraction it produces is then checked against arithmetic that does not involve a model at all.

Classification is the second use. Given a vendor's history and your chart of accounts, suggesting the right expense account is a good fit for a model and a bad fit for a rule, because the pattern lives in the history rather than in a policy anyone wrote down.

Explanation is the third. A variance report that says which account moved and by how much is arithmetic. A sentence saying the movement is the quarterly rent that fell in a different month this year is worth reading, and worth having a person confirm.

Where it must be kept out

Arithmetic, balancing, VAT computation and matching are all deterministic code here. That is not caution for its own sake. A model that is right most of the time produces a ledger that is wrong occasionally and undetectably, and there is no worse outcome for a set of books than one you cannot tell is broken.

The same reasoning drives the approval gate. Proposals from the agents go through deterministic validation, become drafts, and wait. Nothing an agent produces is trusted enough to post.

• Every balance and every VAT figure is computed in plain code, not by a model
• Every proposal is validated before a human is asked to look at it
• Every entry carries the source documents it came from
• Every agent decision is recorded, so a wrong classification can be traced

What it costs against what you pay now

An outsourced bookkeeper in Saudi Arabia runs roughly SAR 1,500 to 7,000 a month depending on transaction volume, and that figure leaves out the owner's own hours. ProCul is SAR 1,499 a month. The comparison worth doing is not the invoice against the invoice, it is the invoice plus your hours against the invoice plus about two hours a month answering questions.