Financial directors prefer a reliable and predictable AI to an autonomous but unpredictable AI.
The development of the AI agents is fast. Today, only 6 % of financial officials use them, but 38 % intend to do it within a year, according to Wolters Kluwer (May 2025). However, the more the autonomy progresses, the more the question of control becomes crucial, especially in an area where each decision engages responsibility, compliance and precision.
The example of the Project sells with Claude, the agent IA of Anthropic has demonstrated it: autonomous, brilliant … and disastrous.
An autonomous AI… and uncontrollable
At the end of June 2025, Anthropic launched a daring experiment. His model of artificial intelligence, Claude Sonnet 3.7, was placed at the head of an online micro-shop, with a mission: to manage all commercial operations alone. Price fixing, discounts, control of orders, customer relations … Claude had carte blanche.
The result was as fascinating as they are worrying: awarded to all-VA, prices set below the cost cost, invented payment methods, and in the end, a store led to bankruptcy by excess of creativity.
This demonstration highlights an essential question for businesses: what happens when the autonomy of AI is not supervised? In a commercial universe, this can result in losses. But in a financial direction, the impact can be much more serious: accounting errors, regulatory breaches, and loss of traceability.
No improvisation in finance
AI is now invited to the heart of business processes, including the most sensitive. But the question is no longer whether it can be useful, it is whether it can be reliable, especially when it affects areas as critical as finance.
Experience has shown how inventive AI can be, but also unfit to integrate fundamental economic logics. Claude did not act badly: he simply acted without awareness of the result. He automated … without understanding what he was doing.
What financial directors are looking for today is not an AI that takes initiatives. They don’t need a “brilliant” agent who promises to revolutionize everything. They need a reliable, repeatable, silent and above all flawless assistant. AI must be able to accomplish well -defined tasks, with constant precision, and have perfectly predictable behavior. In other words, AI becomes “boring”, but its reliability is unassailable.
Determinism as a guarantee of confidence
According to expert Jiquan Ngiams, co-founder and CEO of Lutra AI, “reliable and deterministic systems will be much more desirable than tape-to-the-eye agents that aim to do everything”. Thus, the challenge is not to imitate human intuition, but to reproduce a business logic without error.
It is in this perspective that AI finds its true value: when it automates the tedious tasks which mobilize human time unnecessarily, while respecting the rules.
Treatment of supplier invoices, for example, can be accelerated by agents capable of automatically generating invoices and predicting the right accounting codes with precision greater than 90 %. This kind of solution is not spectacular, of course, but it works. And this is precisely what makes her precious.
Innovate, yes. But under control.
It is therefore not a question of refusing innovation, but of supervising it. The financial agent of the future will not be a creative genius but a disciplined operator, respecting the processes and leaving a clear trace of his decisions.
The project sells reminds us of how the AI can wander when it is delivered to itself. And how urgent it is, for companies, and even more in the finance function, to favor supervised artificial intelligence, built not to impress, but in which we can trust.
