IFRS 18 Pulls Adjusted EBITDA Into the Audited Accounts. Your AI Can't Improvise the Math

For years, “adjusted EBITDA” lived in the safe part of the report. It sat in the investor deck, the earnings call script, the lender covenant pack. It was management’s number, defined by management, explained in a footnote nobody initialed. The external auditor signed the statutory accounts; the adjusted figure floated alongside, useful and unaudited.

IFRS 18 ends that arrangement. The standard is effective for annual reporting periods beginning on or after 1 January 2027, with comparatives, which means your 2026 figures get restated and presented under the new rules. For the first time, management-defined performance measures, the category that captures adjusted EBITDA, adjusted operating profit, underlying earnings, and their cousins, are brought inside the financial statements and made subject to audit. The reconciliation from the MPM back to the most directly comparable IFRS subtotal becomes a disclosure note. The auditor tests it like any other note.

That single change rewires how a finance function should think about calculated figures. And it lands at an awkward moment, because a lot of teams are now wiring large language models into financial reporting.

What actually changes in the numbers

Two things, and they compound.

First, the face of the income statement. IFRS 18 classifies income and expenses into operating, investing, financing, income taxes, and discontinued operations, and mandates two new defined subtotals: operating profit, and profit before financing and income taxes. These aren’t cosmetic. The classification of an item can depend on the entity’s main business activity, so a financing cost for an industrial group and a financing cost for a bank don’t land in the same place. The subtotals are defined, ordered, and auditable.

Second, MPMs. If you publish adjusted EBITDA in a press release or an analyst call, IFRS 18 says you must disclose it in the notes, define it, explain why it’s useful, and reconcile it to the nearest IFRS subtotal. Every adjustment line, the one-off restructuring charge you add back, the impairment you strip out, the share-based payment you exclude, has to be reconciled and consistent period over period.

So you now have a chain: ledger detail rolls into IFRS 18 categories, categories roll into defined subtotals, subtotals reconcile to your adjusted measures. Break any link and the audit finds it.

Where AI helps, and where it quietly hurts

A language model is genuinely good at parts of this. Drafting the MPM narrative. Explaining why the standard reclassifies an item. Summarizing the change for the audit committee. Spotting that you described an adjustment differently in Q1 than in Q3. Use it there.

The problem starts the moment you let the model compute the figure itself. Ask an LLM to “calculate adjusted EBITDA from these accounts” and it will produce a number. It will look right. It may even be right. But it was generated by predicting tokens, not by executing a defined formula against reconciled source data. Run the same prompt twice and the add-back list can drift. Change the phrasing and the operating-profit subtotal shifts by a reclassification the model decided to make on its own. None of that is acceptable when the output is a number an auditor will tie out to the ledger.

Here is the uncomfortable part for CFOs leaning into automation: the more fluent the model, the more dangerous the wrong answer, because it arrives with confident prose attached. An MPM reconciliation that’s off by one add-back, presented in clean English, is harder to catch than an obvious error.

The fix is boring on purpose: separate retrieval from calculation

The reliable pattern is to stop the model from doing arithmetic and instead let it call a deterministic engine that does. This is the architecture Rexfin is built around.

Connect the source, your QuickBooks, Xero, NetSuite, Sage, SAP, or Oracle ledger, your warehouse, or uploaded statements, and Rexfin builds one reconciled financial model that ties out to the ledger. That model is the single source of truth. Adjusted EBITDA isn’t a phrase the AI interprets; it’s a defined calculation in the engine, with a named formula, a fixed adjustment list, and the same logic applied every period.

When someone asks a question, the AI retrieves the relevant figures and runs the math through the deterministic layer, not through the model’s own token prediction. The number comes back identical every time you ask. The IFRS 18 operating-profit subtotal is computed from classified ledger items by a rule you can read, version, and hand to the auditor. The MPM reconciliation is generated from the same engine that produced both ends of it, so the bridge always foots. And because every figure traces to source, the audit question, “show me how this number was built,” has a one-click answer instead of a week of reverse-engineering spreadsheets.

This matters most in the GCC right now. Groups preparing FY2025 audited statements during 2026 are running straight into the IFRS 18 comparatives, and many are doing it across multiple entities and SPVs where a single inconsistent add-back multiplies across the consolidation. The same discipline that satisfies the auditor, every number tied back to source, is the discipline that makes audited statements traceable for AI in the first place.

What to do before the 2026 comparatives lock

A short, unglamorous checklist:

• Write down every MPM as a formula, not a description. “Adjusted EBITDA = operating profit + depreciation + amortization + [named, fixed add-backs].” If you can’t express it as a rule, you can’t audit it and you certainly can’t let software reproduce it.
• Map your ledger to the five IFRS 18 categories now, against 2026 actuals, so the operating and pre-financing subtotals are computed by rule rather than judged line by line at year-end.
• Decide where the model is allowed to think and where it isn’t. Narrative and review: yes. Producing the figure that hits the note: no. The calculation belongs to a deterministic engine.
• Keep the trace. Every published number should reconcile to a subtotal, every subtotal to classified ledger items, every item to source.

The standard isn’t asking for heroics. It’s asking for consistency that holds up under audit, which is exactly the thing free-form LLM math can’t promise. Adjusted EBITDA is now a controlled number. Treat it like one: define it, compute it deterministically, and let the AI explain it rather than invent it.

If you’re sizing up how IFRS 18 changes your reporting stack, book a demo and we’ll show you the same number coming back the same way, every time, with the trace to the ledger sitting right behind it.

Want the wider picture first? The trusted numbers layer for AI in finance pillar covers consolidation, tax workpapers, and audit traceability in one place.