At the same cost, AI transforms your financial analysis. The skill of judging it makes the difference.

An artificial-intelligence model can now draft, in a few minutes, a results commentary, a credit note or a company valuation that look right. The cost of producing a plausible financial analysis has fallen practically to zero.

The cost of verifying it has not moved. And that is excellent news for those who know how to read it.

Here is what few executives have yet taken on board. On a constant budget, AI can multiply the efficiency of financial analysis: where producing a note used to take hours, it now takes minutes. The time freed up does not disappear; it shifts to what really creates value, judgment and then action. But this shift only happens on one condition: mastering the fundamentals that allow you to control the machine. Without that base, AI does not save you time. It manufactures, faster, well-presented errors.

A concrete example. Producing the monthly results commentary used to take a CFO two to three hours: extracting the figures, formatting them, writing. With AI, that part drops to about twenty minutes of production and review. On the same budget, more than two hours remain. Everything depends on how they are used: reinvesting them in action. Spotting that customer payment terms have slipped from 45 to 62 days, deciding to renegotiate deadlines with the two largest accounts, quantifying the effect on cash. AI did not create that value: it freed up the time to go and find it where it lives, in action.

Value is therefore born neither from the tool alone nor from the human alone: it is born from their meeting. An AI that produces, a human skill that decides whether or not to trust it. That is exactly the skill we build at Sagora through training and advisory. And it is the meaning of our vision: finance by humans and for humans. In other words: finance that only makes sense when it informs real decisions and serves the people who carry them.

For an executive or a finance manager, the rare skill is no longer producing the analysis: it is knowing which question to ask, recognising a wrong figure and spotting the frontier beyond which the tool breaks down. Three sets of data put that diagnosis on an objective footing. Here they are, with what they mean concretely for your decisions.

Understand: real capabilities, narrower than the promise

Let us start by measuring rather than believing.

Vals AI’s “Finance Agent” benchmark puts the large AI models through 927 questions modelled on the work of a junior financial analyst: searching through hundreds of pages of published accounts, year-on-year adjustments, multi-step calculations. In its May 2026 version, the best model tested tops out at around 52% correct answers. In the most demanding categories, including financial modelling, the best models fall to 23% correct answers.

So the tool is useful. It is not reliable on its own. Above all, its reliability is not uniform: it draws a jagged frontier, strong here, failing there.

That is precisely what “Navigating the Jagged Technological Frontier” (Dell’Acqua et al., 2023) shows, the most-cited field study on the subject, run by Harvard Business School with the Boston Consulting Group among 758 consultants. Inside the perimeter where AI performs: around 25% time saved and clearly higher quality. Outside that perimeter: 19 percentage points less success than the group working without AI. The tool did not merely stop helping; it degraded performance.

The most troubling point of the study fits in one sentence: people delegate most where the tool is weakest. Because the frontier is invisible, trust is poorly calibrated. And it is that poor calibration, far more than the technology itself, that constitutes the risk. Training yourself means, first of all, learning to see that frontier.

Decode: the danger is not the visible error, it is the error that looks right

A crude error gets spotted. A well-presented error gets signed.

In October 2025, Deloitte refunded the final instalment of a contract worth a total of 440,000 Australian dollars concluded with the Australian government, after a dozen non-existent references and a mis-cited court ruling were found in the delivered report. This is not a beginner’s blunder: it is what happens when a generative tool enters the production chain without the control chain being rethought.

The most accurate image is this one.

Nobody would sign a first-week intern’s note with their eyes closed. The real question is why we are tempted to do so as soon as the same text comes out of a machine.

An example from our teaching practice illustrates it well. Asked about the value of a company, AI spontaneously proposes a multiples approach: applying to your EBITDA the coefficient observed in comparable transactions. That is the market’s dominant practice, and that is precisely the problem. Multiples are a market reference point, a language of comparison, at best a consistency check: they do not constitute a valuation method. Two companies with the same EBITDA can differ in value by a factor of two depending on their growth and the quality of their management. The value of a company corresponds to its discounted future cash flows, and the real debate lies in the assumptions: the opportunity cost of capital, normative growth, the sustainability of the flows. When you correct an assumption the tool had set on its own, without saying so, you measure how far the value moves. The tool reproduces dominant practice; it does not choose the correct method for you. A teacher does.

One last element managers still underestimate: responsibility cannot be delegated. The European AI Act (Regulation (EU) 2024/1689) has required since February 2025, through its Article 4, a sufficient level of AI literacy from organisations deploying these systems, and most of its obligations apply from August 2026. Signing a figure produced by a machine is still signing, with all the consequences in terms of responsibility.

Act: three reflexes to install before the next report

What to do with all this, concretely? Three reflexes, in this order.

Consolidate the fundamentals, do not replace them. AI has not changed financial grammar: a cash flow is still a cash flow, an assumption is still an assumption. That is precisely the method that allows you to control the machine. A manager who cannot build an analysis will not be able to verify one, and will end up signing what he has not understood.

Map the frontier by testing it, not by decreeing it. Distinguish what the tool does well and at low risk (data extraction, structuring, formatting, first reading, rewording) from where it is dangerous (choice of assumptions, assessment of data quality, conclusion, recommendation). The best exercise we know: hand a team an analysis note produced by an AI on real accounts, with a double instruction, validate it then break it. You almost always find several errors, including at least one that changes the conclusion. It is the most effective vaccine against poorly calibrated trust.

Ask one question before all the others: where does this figure come from. Every data point used in a decision must have an identifiable source. The confidentiality of company data and the traceability of what was produced are treated as starting constraints, not as a stylistic clause.

Case study: “990 euros, just to press a button?”

A client challenges us, sceptical, about the price of an analysis of his internal accounts: 990 euros. His reasoning fits in one sentence: “All it takes is pressing a button.”

That is exactly the misunderstanding AI installs. The button exists, indeed. What he does not see is everything that happens around it.

I answered him, with a smile: to get light, too, all it takes is pressing a button. And yet we willingly pay for all the technology, the network and the service that make that simple gesture possible. The button is never the product. It is the visible part of work that does have value.

Behind those 990 euros, there is no automatic extraction. There is the development of the analysis, a reading of the figures connected to the reality of his business, the pedagogy that makes the result understandable and debatable and, above all, the action levers: what to decide, in what order, with what expected effect. The deliverable is not a disposable note. It is a financial report that holds up in a board meeting and at a general assembly.

We are not in compliance. We are in analysis: understand, decode, act. Pressing a button produces a text. Understanding what it says, verifying it and turning it into a decision: that is what has a price, because that is what has value. Finance by humans and for humans, down to the lever of actions serving the teams.

The deposit is real, and it is still largely untapped

These precautions are not a case against the tool. The gains are concrete for those who use it within its perimeter: the Harvard-BCG study measures them, around 25% time saved and higher quality where the tool is in its place. A manager who masters this frontier produces faster, explores more scenarios and spends his time where it counts: on judgment.

And that advantage is still up for grabs. According to “The GenAI Divide: State of AI in Business 2025”, a report by Project NANDA, hosted at the MIT Media Lab, around 95% of generative-AI pilot projects run in companies produce no measurable impact on the income statement. And Gartner’s surveys of finance departments put internal skills and data quality at the top of the barriers to adoption, ahead of technology and ahead of budget. In other words: what is missing is not the technology, it is the skill to turn it into value.

It is often said that AI will not replace the finance manager, but that the manager who knows how to use it will replace the one who does not. The saying is only true on one condition, and it changes everything: knowing how to use it does not mean knowing how to make it produce, it means knowing how far to trust it. The shift to make fits in one sentence: moving from “what can the tool do in my place” to “what must I, myself, be able to verify”. That answer is what has value, and it is what will remain true when the tools have changed names.

That is exactly what we work on at Sagora by aligning training and advisory: the fundamentals that allow you to judge, and the tool’s frontier, tested on real cases. At the same cost, that mastery is what decides whether AI saves you time or makes you sign faster.

The Sagora offer, concretely

Putting this mastery at your service is not just an idea. Here are the four ways we make it operational.

• Analysis report with an explanatory call. From €990 excl. VAT. A complete analysis of your accounts filed with the National Bank of Belgium. Above all, a 45-minute call in which we explain the results with clarity and pedagogy: we translate the figures into concrete action levers, so you know what to decide and in what order. You do not receive a report, you leave with a plan.

• Analysis report on internal accounts. The same rigour applied to your internal accounts, as close as possible to management reality. Fine-grained steering rather than a legal-filing snapshot, with the levers to activate first.

• Chosen competitive benchmark. You choose the competitors, we make them readable. You position your performance and pinpoint exactly where to act to close the gap.

• Training and advisory. The fundamentals that allow you to judge and the tool’s frontier, tested on your real cases. You gain the autonomy to turn every analysis into a decision you own, without depending on the tool.

Let’s talk: formations@sagora.eu.

Sources

• Vals AI, “Finance Agent Benchmark” (v2), May 2026: www.vals.ai/benchmarks

• Dell’Acqua, F. et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, Harvard Business School Working Paper 24-013, September 2023: hbs.edu

• International press on Deloitte’s partial refund to the Australian government (DEWR report), October 2025, notably Fortune.

• Regulation (EU) 2024/1689 of 13 June 2024 (Artificial Intelligence Act), Article 4: eur-lex.europa.eu

• Project NANDA (MIT Media Lab), “The GenAI Divide: State of AI in Business 2025”, 2025 report.

• Gartner, “Gartner Survey Shows Finance AI Adoption Remains Steady in 2025”, press release, November 2025.

Sagora Finance, editorial line. Article based on the Sagora training and advisory programmes and on the Sagora Analytics offer (author: Prof. Mathias Schmit).

Disclaimer: This article is for informational and educational purposes. The data cited come from the sources listed above; the client example is anonymised. It does not constitute investment advice.

www.sagora.eu