AI in financial consolidation: more to sign, or more reason to sign? 

Day five of the close, and the exception list is still open. Intercompany logic lives in two people’s heads. Subsidiaries submit late, each in its own format. The auditor asks who changed a figure, and nobody can show it without an investigation. Somewhere in the background, the consolidation system is running out of maintenance. 

Into that week walks AI, offering to prepare more of the work for you. 

The short version: AI in financial consolidation should explain and propose. Rules should do the checking, and you should do the signing off. Before you let any tool near the close, ask two things. Where does its mistake land? Can you still see the number as it was before the AI touched it?

You already sign off on work you didn’t do 

This part never makes it onto a slide. As group controller you sign off on the consolidated number, and that number is built from reporting packages prepared in other countries, on other ledgers, by people who report to someone else. 

At Ammega that’s 72 legal entities on several ERPs. At Zentiva it’s 70 reporting units in more than 30 countries. it’s 70 reporting units in more than 30 countries. 

You’re not flying blind. There’s a group accounting manual, validation rules on every package, a sign-off from each local finance lead and an analytical review on top. All of it still comes down to relying on work you didn’t do. Chasing every warning to the bottom has never fitted into close week. 

Controllers are rewarded for the absence of failure. Nobody sends flowers for a clean close, and one restatement can end a career. That’s a lot of weight on one signature. 

The usual AI pitch gives you more to sign 

A lot of what’s sold for the close works the same way. The AI prepares the reconciliation or the entry, and the accountant moves from preparing to reviewing. 

That may suit an accountant in one entity. A group controller was never the preparer. You’re being handed a second pile of work you didn’t do, stacked on the first. 

It also turns your review into the control. COSO, the body behind the most widely used internal control framework, published guidance on generative AI in February 2026. It says controls “should treat outputs as claims requiring validation, rather than as facts to accept by default”. Where a control relies on AI output, COSO expects “the same evidence standards expected for ICFR”. A signature on an AI-prepared reconciliation doesn’t get you there by itself. 

Busy reviewers miss things, too, and experience doesn’t protect them. In a 2023 study at University Hospital Cologne, radiologists with more than 15 years’ experience read mammograms correctly 82% of the time when the AI’s suggestion was right. When the AI’s suggestion was wrong, they managed 45.5%. These are people trained to doubt. Now add close week. 

Two questions for any AI in the close 

Every AI tool in finance carries the same small print: “AI can make mistakes.” Good. Believe it, and plan for the mistake. 

1. Where does the mistake land? If the AI posts an entry, its error sits in the books until someone catches it. If the AI flags a problem in a package, its error is a wrong flag on a list. You can live with a wrong flag. 

The flag that never fires is the one to worry about. A clean list is not a clean package, so the AI’s silence can’t replace a control you already run. 

2. Can you see the number before the AI touched it? This is the auditor’s question from day five: who changed this figure, what was it before, and who accepted the change? If the agent’s work is mixed into the ledger, you’re back to an investigation. If it’s kept apart, you show the before and the after, and back it out if you need to. 

AI can draft. The draft just has to pass question two.

Don’t make the AI the control 

One distinction matters more to your auditor than any other. A validation rule runs the same way every close, and it can be tested. Generative AI doesn’t work like that. In COSO’s words, “GenAI is probabilistic, not deterministic.” 

So let the rule find the breach and let the AI explain it: which rule broke, the likely accounting cause, the fix. A poor explanation costs you a second look. The breach itself was never in doubt. 

Coverage stays your job. Don’t count on AI to notice the check nobody wrote. The judgement calls stay with people as well: scope of consolidation, impairment, provisions. 

One group got a shorter close and a number it can trace 

Half of finance teams take six days or more to close, according to a benchmark reported by CFO.com, and you know who’s asking for those days back. 

Zentiva was carved out of Sanofi in 2018 and has made ten acquisitions since 2019. It runs IFRS consolidation and planning in Board. Data uploaded by its legal entities is “validated, reconciled, and consolidated”, in that order. 

The monthly close got two days shorter. At the same time Zentiva added reporting layers and transaction-level detail, which Group Reporting Manager Stanislav Drahoš says “significantly improve data traceability, transparency and data quality across the reporting framework”. 

“Board enabled us to shorten the monthly close by two days and deliver more granular financial statements, including detailed cash flow at group, regional, and country level.” 

They didn’t have to choose. The two go together: a break that’s explained on day two isn’t sitting on the exception list on day five. 

Days to close tells you how fast. These tell you whether it held: 

• reporting packages accepted first time 
• top-side adjustments booked after the packages are in 
• intercompany differences still open at cut-off 
• audit adjustments at year-end 

 Where AI belongs in financial consolidation: on the controls 

Three checks sit between the reporting package and the consolidated number. You run all three already. 

• Account mapping. Local accounts map to the group chart of accounts. IFRS 18 applies to periods beginning on or after 1 January 2027, with comparatives restated. If you report on the calendar year, the comparative year is 2026, the one you’re closing now. Profit doesn’t change. Where each line sits does: operating, investing or financing, plus two new subtotals. FX is the awkward one. Exchange differences follow the item that caused them, and most charts of accounts hold FX in a single line. No mapping fixes an account that mixes the two, so that one needs splitting at source. For the rest, an agent proposes the mapping and finance confirms it.that one needs splitting at source. For the rest, an agent proposes the mapping and finance confirms it. 
• Trial balance validation. The rule fires. The agent shows which rule, the computed imbalance, the likely accounting cause and the expected correction. The accountant makes the fix. 
• Intercompany reconciliation. The agent sorts each difference by type, such as a missing or wrong booking, and ranks the biggest gaps. Your intercompany policy still decides who moves. 

In Board, those three checks are the Controller Agent’s job, and it’s live. It runs inside Board’s consolidation product, which Board maintains: a standard close, a release plan and an upgrade path. Finance configures the rules instead of scripting them, so the logic stays when its author leaves. That’s one answer to the two-people’s-heads problem. 

Source data, adjustments and postings sit in separate layers. What the agent proposes is kept apart from what finance entered, so the before and the after stay visible. Every figure traces back. 

Start with one check 

You don’t have to bet the close on this. Gartner’s advice to CFOs in August 2026 was to pilot agents on “a low risk, contained workflow where errors are visible and reversible”, and to judge the pilot on “governance readiness, not just autonomy or ROI”. Visible and reversible: those are the two questions again. 

• Pick one check you already run. Trial balance validation is a good first one, because the rule does the checking and the agent only explains. 
• Run the agent alongside your team for two closes. Change nothing else. 
• Keep a log of every wrong flag, every miss and every fix that was accepted. 
• Walk your audit team through it before year-end. Auditors dislike surprises more than they dislike AI. 

What to say when the CFO asks about AI in the close 

They will ask. This answer holds up: 

“We started where a mistake lands on a list. Rules check every reporting package, and agents explain what broke. Anything an agent drafts sits apart from what we entered, with the before and after on record. Nothing reaches the consolidated number without one of us accepting it.” 

You’d be in good company. Error and anomaly detection is already the third most common use of AI in finance, at 34% in Gartner’s 2025 survey of 183 CFOs and senior finance leaders. 

The close you can defend 

The close doesn’t end at filing. In Board, the consolidated result is the baseline the next forecast starts from, so the number you signed off is the number the plan starts from. Price, hiring, investment and cash all get decided from it. 

That’s the close worth wanting: one your auditor can follow, and your plan can start from. 

Watch the three Controller Agent demos 

FAQ