Five AI workflows every MENA finance team should automate in 2026 — and the prompts to start today

Every "AI for finance" pitch shows the same slide — a glowing robot closing your books while you sleep. It's the wrong picture. AI doesn't close your books, and it doesn't own a single number on your return. What it does is take the three or four jobs inside the close that are pure reconciliation and narration — the ones that eat your week and need almost none of your judgment — and hand you a first draft.

So this isn't a think-piece about whether AI will replace finance teams. It's a list. Five workflows a MENA finance manager actually meets every month, with a prompt or a clear technique for each, and an honest note on where the tool stops and you start.

One thread runs through all five: the AI drafts, you decide. Keep that and these are safe. Forget it and you've automated your mistakes.

1. VAT return prep — catch the mismatch before submission

The highest-stakes repetitive job on the calendar, so start here. The pattern: hand the model your invoice listing and the return you're about to file, and ask it to reconcile declared output and input tax back to the underlying invoices — flagging anything that doesn't tie out before it goes to the authority, not after a query comes back.

I wrote a full breakdown of this one separately — a single prompt that reads a whole VAT return and checks its own arithmetic — because the technique is finicky and the failure mode (a confident, fabricated total) is the worst output a tax form can produce. If you automate one thing this year, automate the check, not the filing.

2. A 13-week rolling cash-flow forecast from your trial balance

The thirteen-week forecast is the most useful number a MENA business owner never has on time. Build it once as a prompt and you can re-run it every Monday off a fresh export.

Export your trial balance, your AR aging, and your AP aging. Give the model your known fixed outflows — payroll, rent, loan repayments, VAT due dates — and let it project the rest from payment history.

You are a cash-flow analyst. I will give you three exports — a trial balance,
an AR aging, and an AP aging — plus a list of known fixed outflows with their dates.

Build a 13-week rolling cash-flow forecast, one column per week, starting this Monday.

Rules:
- Opening cash = the bank/cash balance from the trial balance. State which line you used.
- Time receivables by their aging bucket and our stated average collection pattern;
  time payables by due date. Do NOT assume everything pays on time — show your timing logic.
- Add the fixed outflows on their given dates exactly as provided.
- NEVER invent a figure. If an input is missing or ambiguous, leave the cell blank,
  list it under "Assumptions I could not make," and continue.
- Output: (A) the weekly forecast as a table with a running closing balance,
  (B) the three weeks where cash is tightest, (C) every assumption you made, listed.

The output is a draft, not a promise. You own the collection assumptions and you sanity-check the tight weeks by eye — but you start from a built forecast instead of a blank sheet. In my experience that's the difference between a forecast that gets refreshed weekly and one that gets built once and abandoned.

3. Month-end commentary from a variance table

Management accounts go out late for one boring reason: someone has to write the commentary, and writing the same variance paragraph twelve times a year is nobody's favourite hour. This is the single best use of a language model in finance, because narration is exactly what it's built for.

Feed it the numbers table — actuals, budget, prior period, variances — and let it draft the prose. Then you correct the why, which is the only part that needed you.

Here is a management-accounts variance table: actuals vs budget vs prior month,
with variance columns in value and %.

Write the board commentary as 4—6 short paragraphs. For each material variance
(over the threshold I give below), state the direction, the size, and which
line drove it — quoting only numbers present in the table. Do NOT speculate on
the cause; where a cause isn't in the data, write "[reason — to confirm]" so I
can fill it. Plain, direct English. No filler, no "in today's environment."

Materiality threshold: [e.g. —5% and —EGP 50,000].

It tells you what the numbers say. You supply why — the lost contract, the one-off, the timing slip. That division of labour is the whole point: the model is fast at description and useless at cause, so let it own the first and never the second.

4. AP reconciliation — supplier statement against the ledger

Matching a supplier's statement to your AP ledger is tedious in exactly the way machines are good at: line up by amount, date, and reference, then surface what doesn't match. The time saving here is among the largest of the five, because the work is almost entirely pattern-matching and almost none of it is judgment.

Give the model the supplier statement and your AP ledger for that vendor and ask it to do three things: match what reconciles, separate genuine timing differences (an invoice you've booked that they haven't, or vice versa) from real discrepancies, and list the unmatched items as a worklist. The teams I work with report this turns a half-day of statement reconciliation into a short review of an already-sorted exception list.

The trap is the same as everywhere else: a model will happily declare a match that's off by a digit if you let it be "helpful." Tell it to treat any amount that isn't an exact match as unmatched, and to never adjust a figure to force a tie.

5. Intercompany eliminations across entities

If you consolidate a MENA group — entities in Egypt, the UAE, Saudi, maybe a free-zone holdco — intercompany eliminations are the quiet monster at period end. Entity A's receivable from Entity B should equal B's payable to A. It rarely does, and chasing why is slow.

Point the model at the intercompany schedules from each entity and ask it to pair them up: this receivable to that payable, flag every pair that doesn't reconcile, and split the causes it can see — different posting dates, an FX rate applied on one side, an invoice booked in one entity and not the other. It produces the mismatch list. You decide the elimination, because that's an accounting judgment and a currency-translation call, not a lookup.

This is the workflow with the messiest inputs, so it's the one where "never guess" matters most. Multiple charts of accounts, multiple currencies, multiple posting conventions — point AI at that mess without guardrails and it scales the mess. With them, it just finds the breaks faster than you would by hand.

---

What none of these do

Notice what every workflow above has in common. The AI reconciles, drafts, and flags. It never signs, never assumes, never decides.

• It doesn't own an assumption. The collection pattern, the materiality threshold, the elimination call — those are yours. The model executes them; it can't choose them.
• It will fabricate a total if you let it. Every prompt here forbids guessing and forces a blank-and-flag instead. That line isn't optional — it's what separates a tool you can trust from one that quietly invents.
• Data quality sets the ceiling. A bad export is bad output, faster. None of this fixes a messy ledger; it just exposes the mess sooner.
• It is not your reviewer. It makes review faster and better-aimed. It doesn't remove it.

The teams pulling ahead in 2026 aren't the ones with the cleverest model. They're the ones who worked out which four jobs to hand over — and which four to never let go of. Which side of that line does your close cycle sit on right now?