AI in Finance for the GCC: A Trusted Numbers Layer

A finance director in Riyadh told me his team had three AI tools approved, budgeted, and live. None of them produced a number he would put in front of his board. He used them to draft emails.

That captures the strange position Gulf finance is in right now. The region pours money into AI faster than almost anywhere on earth. McKinsey put GCC enterprise AI adoption around 84% in 2025, up from roughly 62% two years earlier. Sovereign programs in Saudi Arabia and the UAE treat AI as national infrastructure, not an IT line item. The ambition is real and it is funded.

And yet the value isn’t showing up in the finance function. Across GCC C-suite surveys, the pattern repeats: most organizations have piloted generative AI, a small minority can point to measurable financial impact. The blocker almost never turns out to be the model. It’s the data the model is standing on, and whether anyone can defend the number it produced.

For a CFO in Dubai or Jeddah, that last part isn’t academic. You are operating inside ZATCA clearance, UAE e-invoicing, SAMA and CBUAE expectations, and PDPL data rules. A confident-but-wrong figure isn’t just embarrassing. It’s a compliance exposure.

Why Gulf finance adopted AI fast and trusts it slowly

There’s a gap between two things that should move together: how impressive AI looks in a demo, and how reliable it is on your actual ledger.

The demo is genuinely good. Ask a model to summarize a board pack and it reads like a sharp analyst wrote it. The trouble starts when you ask it to compute. Large language models don’t calculate. They predict the next token. Numbers get broken into fragments, and the model pattern-matches its way to something that looks like an answer. On clean, isolated arithmetic that often works. On a real consolidation, across entities, currencies, and a chart of accounts that changed mid-year, it produces a figure that sounds authoritative and ties out to nothing.

We pulled this apart in why AI gets financial math wrong. The short version: the failure is structural, not a bug a bigger model will fix.

Stack that on top of the data reality in most Gulf groups. Multiple entities. QuickBooks in one subsidiary, a localized ERP in another, spreadsheets holding the parts nobody migrated. Arabic and English side by side. Eastern and Western numerals. Until those sources agree on a single reconciled number, the most capable model in the world is just summarizing disagreement, fluently.

Compliance makes “show your work” non-negotiable

Outside the region, a wrong AI number is a management problem. In the GCC, the regulatory floor is higher, and it’s rising fast.

Take Saudi Arabia. ZATCA’s Fatoorah platform processed billions of e-invoices in 2025, a sharp jump year over year, and Phase 2 integration waves keep pulling more companies in by revenue band, with deadlines running through 2026. Invoice data has to be stored inside the Kingdom. In the UAE, the e-invoicing framework and Cabinet-level penalties make accurate, source-traceable records a legal obligation, not best practice.

Then there’s how regulators expect AI itself to behave. SAMA and CBUAE guidance lean hard on explainability, human oversight, and audit rights over AI-driven decisions. The CBUAE’s AI guidance goes as far as expecting institutions to retain the ability to override and shut down a system.

Put those together and the requirement is simple to state, hard to fake: when AI reports a number, you have to be able to trace it to source, show the calculation, and prove it wasn’t quietly altered. A black box that emits a figure with no lineage fails that test before the conversation about accuracy even begins. We go deeper on this in building audit trails for AI in finance.

The missing piece is a reconciled numbers layer

Most teams reach for a better chatbot. The actual gap sits one layer down.

Rexfin builds that layer. It connects to your accounting and financial-data sources, QuickBooks, Xero, NetSuite, Sage, your warehouse, or uploaded statements, and reconciles them into one financial model that ties out to the ledger. A single source of truth, not a second copy that drifts. You can see how the reconciled model is built and which systems it connects to.

The mechanism that earns CFO trust is the split between retrieval and math:

• AI retrieves figures from the reconciled model instead of guessing them.
• Calculations run through a deterministic engine, not the language model. The same question returns the same answer, every time.
• What-if scenarios run against the verified baseline, so a forecast is reproducible rather than a one-off generation.
• Every output traces back to source, ready for ZATCA, SAMA, CBUAE, or your external auditor.

The language model does what it’s genuinely good at: understanding the question, framing the answer. It never gets to invent the number. That division of labor is the whole point of a trustworthy modeling layer.

Why does this matter more in the Gulf than almost anywhere? Because of how Gulf finance teams are structured. Family-group conglomerates, free-zone entities, and fast-acquired subsidiaries mean a single CFO often sits over a dozen books that were never designed to consolidate cleanly. The intercompany eliminations alone can take a controller days. When an AI tool answers a group-level question without reconciling those books first, it isn’t wrong by a rounding error. It’s wrong because it never saw the real number to begin with. Reconciliation isn’t a nice-to-have on top of the AI. It’s the precondition for the AI being worth turning on.

I’ll be honest about the limits. This doesn’t make AI omniscient, and it won’t rescue source data that’s wrong before it arrives. It moves the trust problem to where you can actually manage it: reconciliation and provenance, instead of crossing your fingers on a black box.

What this looks like across the Gulf

The reliability problem is shared across the region. The texture isn’t.

Compliance has its own shape per market: clearance models, e-invoicing deadlines, retention rules, and storage-inside-the-Kingdom requirements all land differently in Riyadh than in Abu Dhabi. We unpack the specifics in ZATCA, e-invoicing, and audit-ready models.

Language is its own engineering problem. Arabic-first models like Jais, Falcon, and ALLAM have closed the gap on Arabic financial reasoning, but mixed numerals and IFRS-versus-AAOIFI code-switching still trip them up. A capable Arabic model still needs a deterministic layer underneath to do the actual math, which is the argument in Arabic-first finance AI.

And data residency has become a board-level question. With HUMAIN, Qatar’s national AI push, PDPL localization, and explicit audit rights in CBUAE guidance, where your reconciled financial data lives is now a governance decision, not an infrastructure detail. We cover keeping figures inside the Gulf in sovereign AI, sovereign numbers. The starting point for all of it is the trust problem itself: why GCC CFOs don’t trust AI with their numbers.

The order that actually works

The instinct across the region is to automate first and trust later, betting the next model release closes the gap. It doesn’t. A faster engine on bad fuel just reaches the wrong number sooner, and at machine speed that’s worse, not better.

Trust first. Build one reconciled model your numbers tie out to. Make every figure traceable to source and every calculation deterministic. Then let AI retrieve and compute against it. Do it in that order and AI stops being the tool you use to draft emails and becomes the one you put in front of your board.

If you’re a Gulf CFO who wants AI on your numbers without betting your audit on a black box, book a demo and we’ll reconcile a slice of your real data so you can see a traced number end to end.