# Why I'm building an agent-native ledger

> Why the monthly close is the bigger problem for Saudi companies, what an agent-native ledger means in plain words, and what I have and have not proven.

10 Oct 2026, by Ali Asif.

I spent three years at PwC in Al Khobar, and what stuck with me wasn't the sign-offs. It was how much human effort sits behind a trial balance.

In many companies the books get done once a month, by hand: bank lines pulled into a spreadsheet, matched to invoices by eye, a transfer nobody remembers, journals typed at the end. The books aren't wrong because anyone is careless. They're wrong because the work happens every thirty days, after memory has faded.

I'm a CPA and an ACCA member, so I can say this without it sounding like a pitch: most of that month-end effort isn't judgment. It's matching. A payout lands in the bank as one number, and behind it sit forty orders and a handful of fees. Finding which is which is arithmetic. The judgment is the small remainder, like classifying an odd payment or deciding whether an accrual is fair, and that part deserves a person's full attention.

## Why Saudi Arabia makes it sharper

ZATCA e-invoicing is rolling out to businesses in waves, with each wave given a date. Every invoice now has to be structured, chained and reported in a defined format. A company that used to get away with a loose ledger now has a regulator-shaped reason to keep it tidy all month. Zakat and VAT returns sit on top of the same books. If the underlying ledger is a month behind, everything built on it is too.

That's the gap I'm trying to build into. The US has a new generation of AI-native ledgers, and Rillet is the one I point to when people ask what I mean. Saudi companies run on different systems, a different tax regime and Arabic documents, so the product has to be made for them from the ground up.

## What "agent-native" means

An old-style accounting system waits for a person to enter things. An agent-native one assumes software does the daily legwork and a person reviews it.

In ProCul, the accounting product I'm building through Procul Solutions, that looks like this. It reads a company's existing books, in read-only mode, and mirrors them into its own double-entry ledger kept separate from the client's. Agents then match settlements and bank lines to the invoices behind them, draft the journal entries, prepare the VAT return and assemble the monthly close.

Two design choices matter more than any feature. First, the arithmetic is plain code, with money held as whole halalas so a float never touches a ledger. A model only reads documents and suggests categories, and the engine runs again after the agents to check them. Second, nothing posts on its own. Every entry sits as a draft until an accountant approves it. I've kept auto-posting off for the first year on purpose, and it would only switch on once the measured accuracy clears a very high bar.

ProCul is not an accounting firm. It doesn't audit, sign or attest. The person who verifies the numbers is the company's own accountant.

## What exists today

The ledger engine, the matching, the schedules (fixed assets, prepaids, loans, leases, revenue), a ZATCA e-invoicing module and connectors for Zoho Books, Xero and Odoo are built. There are 1,238 automated tests and they pass.

The most useful check so far used a synthetic company I built, a Saudi trading and contracting business with 2,283 source documents and 20 planted errors. I ran the product against the company's own reported trial balance. It found the SAR 1,677,448 difference to the halala and flagged the twelve account slots those errors touched. Nothing more, nothing less.

I want to describe that at its actual size. It's a test against data I made. It tells me the engine behaves, not that a real finance team will trust it.

## What I haven't proven

No company is paying for ProCul. I have no customers, no revenue and no pilot I'd describe as one. The pricing is still a hypothesis.

The rule summaries the product uses for IFRS, VAT and Zakat are written in my own words with citations, and a licensed accountant hasn't yet reviewed them. That has to happen before a client relies on them. A couple of the connectors, for Wafeq and Qoyod, still have details I haven't verified against live accounts, and they stay switched off until I have.

I also don't know yet whether owner-led Saudi companies will give a new ledger the trust a ledger needs. A test suite doesn't earn that trust. A few months of clean closes for real companies will.

## What I'm doing next

The plan is to put it in front of real businesses, starting with a free read of their books, and see where the findings land. If a finance person looks at what we found and says it's right, that's the signal I'm after. If they shrug, I'd rather learn it in the autumn than in a year.

I'll write up what I find, including what goes badly.

Canonical: https://heyaliasif.com/writing/2026-10-13-moves-why-im-building-an-agent-native-ledger