Few years ago, monthly close with manual reconciliation across a couple of ERPs and a board pack landing a week after period end was a defensible setup. The next tier of capability, a real-time consolidated view with drill from any KPI down to a source posting in any subsidiary's ledger, cost a multi-year IT project. Most groups did not buy it, they were right not to because ROI was difficult to justify.

That trade-off has changed, and I do not think the change has been priced in.

What AI changed is not what most people assume. The real-time consolidated view was always a capability. The bottleneck was building it, not running it. AI made building it dramatically cheaper. Account mapping across mismatched charts of accounts is a classification problem, and AI handles those well. Detecting intercompany flows across entities is pattern recognition, and AI handles those well too. The work that used to require a standardization project across every subsidiary now needs a controller approving suggested mappings for about twenty minutes a week.

The cost of bridging dropped. The bar moved. Most of the CFOs I have talked to since starting at Corvenia have not checked it.

What this looks like in practice. Two groups, same shape. Twenty subsidiaries each. Three or four ERPs in the mix. Tripletex on the Norwegian operations, Business Central on the German one, Visma Net for the company acquired in Q1.

Group A still closes monthly, reconciles by hand, presents the board pack the week after.

Group B has a live consolidated view. The CFO drills from any group line to the source posting in any subsidiary. An agent surfaces variances before the controller asks.

Both built the capability starting today. They paid about the same. The difference is what each one did about where the model touches the numbers.

Where the model touches the numbers is the architectural decision underneath, and most new tools make the same mistake. They let one model handle everything in the same pass.

These models are good at classification work, take account mapping across subsidiaries: A subsidiary on Tripletex tags one external services expense one way; another on Fortnox tags the same kind of expense differently. AI maps both into the same group line, the controller approves, and what took six weeks of an analyst now takes twenty minutes.

These models are bad at arithmetic, an intercompany payable of 1,000,000 NOK between two entities needs an eliminating entry of exactly minus 1,000,000 NOK, not something the model is 99% sure is close. A balance sheet that is 99% balanced is not balanced.

Run both jobs through one AI and the math drifts. A small error on one intercompany loan is invisible at the line. Across 40 subsidiaries it compounds into a group EBITDA the auditor finds at year-end, after the board has seen the wrong number twice.

The clean way is three layers. AI on top for the classification work. A deterministic engine in the middle for the arithmetic. Immutable lineage at the bottom that traces every group number back to a source posting in a named subsidiary. Each layer does the one thing it is good at.

I have made this argument before in a different domain. The eight years I spent as CTO of an enterprise AI company building for insurance had the same recurring question in architecture reviews. Where in the stack does the model touch numbers that have to be defensible under regulator scrutiny? The pattern repeats here. The novelty is the AI layer on top. The deterministic core underneath is older than the AI conversation by decades. The mistake most of the new finance AI products are making is treating the new layer as a replacement for the old core instead of an addition to it.

The Group CFO who still runs the manual-monthly setup is not slower at month-end in any way that used to matter. Slower at month-end was tolerable in 2022. The disadvantage now shows up in places that used to be invisible. When the sponsor asks for a 13-week cash flow rolled up across the four ERPs by Friday, you cannot produce it without three weeks of analyst work. When due diligence opens on the entity acquired in Q1, you cannot answer the unit-economics question. When the auditor pulls the lineage thread on a Q3 elimination, you cannot show the source posting it traces back to. The questions are the same questions they always were. The new thing is that the answer is now expected within minutes, because peer groups have it within minutes, because the bar moved.

The blocker is budget shape. The category in the CFO's head is still "consolidation project," the cost model is the multi-year IT engagement they remember from 2022, the 2027 line item is built around a vendor whose pricing assumes that engagement. The new shape, weeks of connection rather than months of replacement, does not fit the budget shape that was already approved.

We built Corvenia for that new shape. Mapping is AI. Eliminations are deterministic. Lineage is immutable. ERPs stay where they are. Qben Infra runs 40+ subsidiaries across four ERPs on this architecture, including Tripletex, Fortnox, PowerOffice Go, and Xledger. Onboarding a new entity is days, not the multi-month integration the IT calendar still assumes.