That number stopped me when I read it, not because it is surprising. Because of what comes after it.

Most of those same teams kept using the tools. They just opened a spreadsheet to verify the outputs before every board meeting.

That is not a trust problem. That is an auditability problem. And it is the difference between a tool your finance team uses and a tool your finance team checks.

Most of the conversation about AI in finance focuses on accuracy. Can the model produce a correct variance explanation? Can it flag anomalies reliably? Can it close faster?

Those are real questions. But they miss the one that actually determines whether AI gets used in a board meeting or stays in the analytics team's sandbox.

The question is not whether the output is correct. The question is whether the person presenting it can defend it when challenged.

A CFO standing in front of investors or a board does not say "the AI said so." They need to show where the number came from, what data it was built on, and what judgments were made along the way. If the system cannot show them that, the output does not go in the pack. It goes into the shadow verification process that runs alongside every AI tool that has not solved this problem.

Why most AI tools cannot show their work

The auditability problem in financial AI is not a feature gap. It is an architecture gap.

When an AI agent reasons across raw ERP data from four different systems, the path from input to output is opaque by design. The agent made probabilistic decisions about what account 4010 in one system maps to in another. It made assumptions about how to handle an intercompany transaction that appears as revenue in one entity and cost in another. Those decisions are baked into the output. They are not traceable.

This is the deeper problem with teaching an agent to "learn the languages itself" rather than normalizing the data first. You get an answer. You cannot show the work. And in finance, showing the work is not optional.

A single hallucinated answer can derail entire workflows. In 2026, enterprises will grapple less with building AI and more with trusting it. - KPMG, Enterprise AI Outlook 2026

The only way to prevent it is to build auditability into the foundation, not bolt it on afterwards.

Three moments where auditability is non-negotiable

There are three specific moments where the ability to trace a number back to its source separates a usable system from an expensive experiment.

The board meeting, where the CFO needs to answer "where does this number come from?" without hesitation.

The audit, where the question is not whether the numbers are right but whether the process that produced them is defensible.

The acquisition, where a new entity's data is being incorporated and the finance team needs to understand exactly how it has been mapped and what assumptions were made.

In all three cases, the requirement is the same. The number came from here. This is how it was calculated. This is who approved the mapping and here is the source transaction.

An architecture decision, not a feature decision

Auditability is not something you add to a financial AI system. It is something you either build in from the start or cannot have at all.

A system that normalizes data through a canonical layer before AI reasoning is applied can provide full lineage. Every output is grounded in a specific, human-approved translation of source data. The AI reasons on structured truth, not on its best interpretation of four incompatible raw sources.

A system that lets the AI interpret raw data directly produces outputs that feel authoritative and cannot be traced. The number exists. The path to it does not.

For a CFO who needs to stand behind their numbers in a room where people will push back, that distinction is everything.

We built Corvenia on the canonical layer architecture precisely because of this. Every number in the consolidated view traces back to a specific transaction in the originating ERP. A controller can drill from group EBITDA to a cost centre, to an entity, to an account code, to the individual posting in Tripletex, Business Central, Fortnox, Microsoft Finance & Operations or whichever system that entity runs. The path is visible at every step.

Every AI-proposed account mapping is reviewed and approved by a controller before it is applied. The system records who approved it and when. If a mapping changes because an ERP config changed, that change is logged. The canonical layer is not a black box. It is a transparent, auditable translation layer where every decision has a human sign-off.

Our customers are consolidating over dozons entities across multiple ERPs in real time. Every number their CFO presents is fully traceable. Not because we added an audit trail feature. Because the architecture produces auditability as a natural consequence of how it works.

AI in finance will keep getting faster, sharper, and more capable. The outputs will improve. The models will get better at understanding financial data.

None of that changes the fundamental question a CFO faces when they walk into a board meeting.

Can you show me where this came from?

If the answer is yes, the output goes in the pack. If the answer is no, the spreadsheet opens.

If you are evaluating AI for your finance function and want to understand what full auditability looks like in practice, we would like to show you.

If your AI tool produced a number your board challenged today, could you trace it back to the source transaction?

‍