Companies audit AI after setting its goals. Philosophy must intervene upstream, before a conception of work or performance becomes a system.
In the spring of 2026, a pontifical text devoted to contemporary transformations formulated with rare clarity what many artificial intelligence (AI) strategies still leave in the shadows. Leo XIV observes that “each design choice expresses a vision of humanity”. The formula directly concerns organizations that entrust an AI with the task of classifying applications, assessing a risk, recommending an action or defining a priority. Behind each of these uses there is a certain idea of work, merit, confidence or performance. Once this concept is translated into data, indicators and decision rules, it acquires the strength of an infrastructure. It becomes quicker to apply and more difficult to discuss. AI governance then often starts too late. It examines the effects of the system after its mission has already been considered obvious.
What the company actually automates
An organization doesn’t just hand over a problem to an AI. She also gives him a way of defining it. In Logic: The Theory of InquiryJohn Dewey shows that a problem takes shape during the investigation which delimits a still indeterminate situation and gives it a direction. Formulating a problem already directs action. This idea becomes decisive when the initial framing can guide thousands of decisions.
Let’s say a company wants to improve its customer service. It can define difficulty as excessive duration of exchanges and ask the AI to reduce the average processing time. A shorter conversation can, however, signal a quick resolution, but also the abandonment of a client or the avoidance of a complex file. The same issue could have been formulated around trust, clarity of offers or lasting resolution. Each framing leads to another intervention. AI can speed up the processing of a request. Defining this demand remains a human responsibility.
The same difficulty appears in recruitment. Identifying the “best candidates” requires defining talent. Should we prioritize immediate suitability for the position, learning potential, the uniqueness of a career path or the ability to transform a function? Each response calls for other data and produces other classifications. The system then makes this definition operational, often with a consistency and speed that makes it seem obvious.
The work of Rachel Etta Rudolph, Elay Shech and Michael Tamir shows that the design of machine learning systems also involves a form ofconceptual engineering. Changing a model can mean changing the way an institution uses social categories. AI thus industrializes less isolated decisions than the concepts that make them possible. An approximate definition, once limited to a few local judgments, can become a rule applied on a large scale. A metric chosen for its availability can gradually transform into an official definition of success.
It is here that philosophy acquires its own role. The technique allows us to understand the system, its capabilities and its limits. Management organizes their integration into work. Ethics examines rights, responsibilities and possible harms. Philosophy comes in much earlier. It clarifies the concepts, compares the purposes and makes explicit the conception of the person that the project is preparing to institutionalize. It prevents a managerial preference, a professional convention or a convenient indicator from too quickly becoming a calculated truth.
Establish philosophical due diligence
To fulfill this function, philosophy must leave the register of commentary and enter the decision-making process. Any significant AI project should be subject to philosophical due diligence before validating the use case. The purpose of this procedure would be to verify that the problem, concepts and purposes entrusted to the system are sufficiently solid to be automated. It should precede the feasibility study and appear in the file submitted to the body responsible for authorizing the project.
The first examination would focus on the formulation of the problem. What situation is the company really trying to transform? Why is AI a relevant answer? What other formulations have been studied? A drop in loyalty can be interpreted as a lack of personalization, a degradation of service, a loss of trust or a value proposition that has become less relevant. As long as these assumptions remain confounded, automation risks hastening an answer without clarifying the question.
The second review would concern the central concept. Which term will be converted into a score, ranking or recommendation? How does the organization define quality, risk, satisfaction or merit? What competing definitions could be reasonably defended? A metric represents a purpose without ever exhausting it. Philosophical due diligence therefore requires the company to distinguish what is easily measured from what really has value.
The third review would focus on assets in tension. Speeding up a process can reduce attention to exceptions. Standardizing a decision can enhance consistency while diminishing the value of situated judgment. Personalizing an offer can increase its relevance while locking the user into the continuity of their past behavior. Philosophy does not resolve these tensions for decision-makers. It makes them visible early enough so that trade-offs can be made before becoming technical routines.
The final consideration would be contestability. Who can challenge the initial definition of the problem? Will the people concerned only be able to report an error, or also discuss the categories that organize the decision? Complete governance must be able to review the mission entrusted to AI, and not just correct its results. Responsibility relates as much to the quality of execution as to the legitimacy of the objective pursued.
Philosophy improves strategy
This clarification work does not slow down the strategy. It prevents it from transforming a fragile intuition too quickly into a technological investment. A request for automation can mask a lack of priority, a conflict between several objectives or a deterioration of the organization. By submitting the initial framing to discussion, the philosophy sometimes reveals that the expected response is less a model than a transformation of the work, the service or the relationship with the client.
This approach also improves the quality of the indicators. Organizations tend to confuse what is easily quantifiable with what really matters. As soon as an objective is converted into a metric, there is a risk of this metric becoming the objective itself. Philosophical due diligence keeps open the distinction between the sought-after value and its numerical approximation.
Finally, it protects the diversity of interpretations. Lisa Messeri and Molly Crockett have shown that AI tools can produce illusions of understanding and foster intellectual monocultures. An organization can generate more analytics while exploring fewer ways to understand its business. The fluidity of the responses then creates an impression of closure. Philosophy reintroduces alternatives. She asks what hypothesis has become invisible, what concept artificially brings together different realities and what purpose has been discarded before being truly examined.
A study carried out among 319 professionals also shows that the use of generative AI shifts the critical effort towards the verification and integration of responses. This move may be useful, but it leaves a deeper question unanswered. Is the proposed framework worth accepting? An answer may be factually accurate and serve a poor problem. Philosophy allows us to control what precedes the verification of facts, namely the construction of the question, the definition of concepts and the justification of the purpose.
Train in a fourth skill
Philosophy should therefore not be added as a peripheral general culture. It must become a practical skill. Know how to problematize a request, distinguish between several definitions of the same concept, explain the goods in tension, construct the most solid objection to a decision and justify a purpose. These exercises can be integrated into AI projects, executive training and investment processes.
As models become available to everyone, the advantage shifts to the quality of the missions assigned to them. An organization can have powerful tools and still be stuck in poor concepts. Another will use the same tools with more discernment because it will have better formulated the problem, better defined the value sought and better organized the protest.
Training in AI therefore requires four complementary skills. Technology allows us to understand systems and their limits. Management makes it possible to organize their integration into work. Ethics governs rights, responsibilities and consequences. The philosophy makes it possible to choose the ends, to clarify the concepts and to preserve the possibility of discussing the mission before automating its execution.
AI lacks philosophy when its objectives seem self-evident. Its mastery begins the moment the organization agrees to question them.
