AI in finance fails on data, not models

Most AI pilots in finance stall for the same reason. Not the model, but the data it is given. What the AI should be able to see, from the first order to the journal entry.

Peter EngelFounder & CEO · 4 min read

Almost every finance team I speak to is trying AI right now. ChatGPT for journal entries, Claude for a transaction export, Copilot to summarise a report.

On a clean file, with an experienced controller next to it, those experiments mostly work. The trouble starts when a team moves from trying AI to relying on it. Then it breaks. Quietly, and with a lot of confidence.

My point is simple. The model is rarely the problem. The data you hand it is.

AI does not say "I’m not sure"

Ask an AI tool about your month-end position and you get an answer. A complete, well-structured, plausible answer. Deloitte puts the risk plainly: in finance, a difference of only 0.5% can, in some situations, amount to millions.

A wrong answer from AI does not look like an error. It looks like every other answer. Picture a close run on data from a PSP, the order platform, billing and the ERP. None of them fully agree. The AI fills the gaps with something that looks right.

In a small team, someone senior usually feels that a number is off. At high volume, nobody can keep up, and an estimate quietly becomes the figure.

Why the pilots stall

The research points the same way. McKinsey’s November 2025 survey of finance leaders found that nearly two-thirds of organisations have not yet begun scaling AI across the enterprise.

Gartner says why. Through 2026, it expects organisations to abandon 60% of AI projects that are not supported by AI-ready data. The same report found that 63% of organisations lack the right data management practices for AI, or are not sure they have them.

Most teams are building AI on a foundation they don’t trust themselves.

An export is not the truth

Most experiments run on exports. Pull the data from the ERP, save it as CSV, load it into the tool, ask your question.

A CSV is a snapshot. It does not know where each record came from or which order belongs to which payment. It cannot tell you whether last week’s settlement has reconciled or is still open. Ask an AI to close the books on that, and it reasons over something incomplete. Transactions that were never in the export simply do not exist for it.

A data lake does not solve this either. It stores data, but it does not keep the relationships between orders, payments, settlements and bookings. McKinsey found that 70% of organisations struggle to get data into AI models quickly. The data exists. It is just not structured in a way AI can use.

You cannot prompt your way out of this. It is an infrastructure problem.

What the AI should be able to see, from the first order to the journal entry

You want answers you can sign off on. Then the chain behind the numbers has to exist before anyone opens the tool. In a high-volume business, that chain has six steps.

  1. The data comes in. Orders from the platform, payments, refunds and fees from each PSP, payouts from the bank.
  2. The data is put into one format. Each source has its own IDs and fields, so every event is translated into the same structure.
  3. Every transaction is matched. Order, payment, settlement and payout are linked. A mismatch gets a reason, or stays visible as open.
  4. Balances are calculated. On any day, you can see what each PSP owes you and what is still open.
  5. Period, revenue and VAT rules are applied. Each amount lands in the month and the account it belongs to.
  6. The journal entries are made and posted. Only matched, explained totals go to the ERP.

Give an AI step 6 and it sees the result. Give it steps 1 to 6 and it sees how the result came about. Ask "why is this week’s Adyen payout lower than our sales?" and the answer leads back to the transactions behind it. That is what I mean by one maintained source of truth. Reconciliation happens before anyone asks, and traceability starts upstream of the ERP.

The human check doesn’t scale

Experienced finance managers catch things. They know their numbers well enough to sense when something is off, and they push back.

That instinct is real, but it is not a process. You cannot document it or hand it over. And at tens of thousands of transactions a day, across payment methods, currencies and channels, no one’s instinct covers it. The risk is not that AI fails loudly. The risk is that it fails silently, where nobody is in a position to notice.

Keep experimenting, but fix the order

Keep trying AI. The teams learning what it does well in their own context are ahead of the ones waiting.

But an experiment tells you which questions AI answers well. It does not tell you whether your data can carry those answers. Gartner’s 60% is a mistake in the order of the work. Teams put money into AI before the data that makes it trustworthy.

So the first question is not which tool to use. It is whether the AI can see what actually happened. If the honest answer is "an export", the work starts upstream.

The Actuals scan takes 15 minutes and shows where your data first stops matching, before any AI looks at it.

Sources

  • McKinsey, How finance teams are putting AI to work today (November 2025)
  • McKinsey, Charting a path to the data- and AI-driven enterprise of 2030 (September 2024)
  • Gartner, Lack of AI-ready data puts AI projects at risk (February 2025)
  • Deloitte Switzerland, AI doesn’t lie, it hallucinates (December 2025)
In practiceBefore AI touches your finance data: is that data right?

Written by

Peter EngelFounder & CEO