FY2025 Audited Statements Are Due in 2026. Can Your AI Tie Every Number Back to the Audit?

A UAE company with a 31 December 2025 year-end has until 30 September 2026 to file its corporate tax return and settle the liability. Nine months from period-end, no extension by default. The Federal Tax Authority treats filing the return and paying the tax as one obligation, so getting close doesn’t count. By the time that deadline lands, your finance team will have spent months producing audited financial statements that an external auditor has signed, the FTA can inspect, and a board has approved.

Now layer in the question that almost nobody asked when they bought their first finance copilot: when a controller asks an AI tool “what was adjusted operating margin for the Dubai entity in Q3,” and the AI returns 18.4%, where did 18.4% come from?

If the honest answer is “the model inferred it from a few documents in the context window,” you have a problem that gets sharper every quarter.

The audit doesn’t care how confident the model sounds

Generative models are built to produce fluent, plausible output. That is exactly the wrong instinct for a number that has to survive an audit. A large language model asked for a figure will return a figure. It may be right. It may be off by a rounding convention, a period boundary, an intercompany elimination it never saw, or a revenue line it half-remembered from a PDF you pasted three prompts ago. The output reads identically either way. There is no visible seam between “retrieved from the ledger” and “guessed with conviction.”

An auditor’s working assumption is the opposite. Every material number in the statements should tie back to the trial balance, and the trial balance should tie back to source. That’s the chain of evidence the whole exercise rests on. A figure that can’t be traced isn’t a minor weakness in the file; it’s an unsupported assertion. Auditors flag those. The FTA, reviewing a return within its assessment window, can ask for the same support.

So the standard for an AI number that touches the audited accounts isn’t “usually accurate.” It’s “traceable to the same reconciled source the auditor relied on.” That is a much higher bar, and most AI-over-finance setups don’t clear it.

Why “feed the statements to the model” quietly fails

The common pattern is to drop the management accounts, a few exports, and last year’s audited PDF into a model and let it reason over the pile. It demos well. It breaks in three predictable places.

First, the figures in those documents rarely agree with each other. The management pack shows one revenue number, the ledger export shows another after a late adjustment, and the prior-year statements use a presentation that has since changed. The model has no principled way to decide which is canonical. It will pick one, or blend them, and won’t tell you it did.

Second, the math happens inside the model. Ask for a margin, a growth rate, or a covenant ratio and the LLM performs the arithmetic token by token. That is non-deterministic by construction. Run the same prompt twice and you can get two answers. No auditor accepts a control that produces different results on identical inputs.

Third, there’s no lineage. Even when the answer is correct, you can’t show why. You can’t point at the journal entries behind it, the entity scope, the period, the FX rate applied. When the auditor asks, you’re back to manual reconstruction, which is the work the tool was supposed to remove.

What “tie back to the audit” actually requires

A number that belongs in an audit file has to satisfy a few non-negotiables. It must come from a single reconciled model rather than a scatter of conflicting exports. It must be computed deterministically, so the same question always returns the same value. And it must carry its lineage, so anyone can walk from the answer down to the source entries.

This is the gap Rexfin is built to close. We connect your accounting and financial-data platforms directly, QuickBooks, Xero, NetSuite, Sage, SAP, Oracle, a data warehouse, or you upload statements when a direct connection isn’t available. From those sources we build one reconciled financial model that ties out to the ledger. That model is the single source of truth, not a copy the AI paraphrases.

When AI sits on top of it, the division of labour is strict. The model retrieves figures from the reconciled layer. Calculations run through a deterministic engine, not the language model, so a margin is a real division of audited inputs and not a generated token. What-if scenarios move from that same base. Every figure in an answer traces back to the entries that produced it. The AI’s job is to find, frame, and explain, never to invent the number.

The contrast matters most under pressure. During the close, with the nine-month clock running, a controller shouldn’t have to ask whether the assistant’s answer will hold up. It either reconciles to the ledger or it doesn’t, and Rexfin only returns answers that do.

The standard is about to tighten

If reconciliation feels like a problem you can defer, the calendar disagrees. IFRS 18 takes effect for annual periods beginning on or after 1 January 2027. Among other changes, it pulls management-defined performance measures, the adjusted EBITDA and adjusted operating profit that companies put in investor decks, into the audited notes, with a required reconciliation to IFRS-defined subtotals. Numbers that used to live in unaudited commentary are moving inside the auditor’s scope. An AI that improvises adjusted figures will be improvising inside the audited accounts. We work through that shift in IFRS 18 puts adjusted EBITDA inside the audited accounts.

For groups, the pressure compounds. A holding structure spanning dozens of entities has to consolidate, eliminate, and present on the same nine-month clock, and AI that can’t keep entity scope straight will produce numbers that look right and reconcile to nothing. That’s the subject of one reconciled model across 40 SPVs.

The takeaway

The UAE’s nine-month deadline isn’t only a filing rule. It’s a forcing function. It puts a hard date on the moment your numbers face an auditor and the FTA, and it exposes any tool that produces figures nobody can trace. The fix isn’t a more careful prompt or a bigger model. It’s architecture: one reconciled source, deterministic math, and lineage on every figure, so the AI a controller leans on is anchored to the same ledger the auditor signs.

If you’re heading into a 2026 filing and want to see what audit-traceable AI actually looks like over your own books, book a demo. We’ll connect to your data, reconcile it, and show you a number, then show you exactly where it came from.