AI lacks context, and companies risk losing theirs

20 October 2026 · @Thibault GOURDIN-SERVENIERE

When Shengxing Yang begins his thesis in 2021, ChatGPT does not yet exist; it will launch a year later. The topic he is offered, how European companies can prepare for AI transformation, concerns something nobody knows how to describe. At the time, people talk about machine learning and specialised systems, imagining what machines might do. “Everyone was saying something, there was no shared vision or common agreement. I thought: I really am free to speak,” he recalls.

His literature review, from the social sciences to writings by futurists worried about employment, yields no common definition. Even today, he observes, the term mainly refers to generative AI, the AI of large language models. From this ambiguity, he developed the idea that runs through his work: “Without a concrete context, you cannot really see what benefits these tools can bring you.”

What the models do not know

The researcher is no sceptic. He recently paid 200 euros for a month’s ChatGPT subscription and says he built an online software application in a day, something he would not have imagined possible without these tools. Of the major model developers, he says: “They have computing resources and substantial investment in data and infrastructure. […] But what escapes them is context.”

Their success, he explains, rests on an earlier investment: the internet and its vast store of public data. The rest remains beyond their reach. In medicine, for example, researchers who do not share their specific data prevent these companies from tackling a precise task seriously.

He tested this in his own field. After incorporating his analytical framework into the tool he developed, he initially thought the consulting profession was doomed: simply apply the framework to a client’s situation to obtain a report. He reconsidered, because without personalisation the result amounts to a well-structured discussion. What remains human, in his view, comes down to taste: personal experience and what makes a life and a career unique.

His own situation illustrates the argument. Living in Paris without speaking French, he considers consulting in France almost closed to him, because language carries context. Machines are no exception: use inaccurate terminology in a prompt and the answer goes off track.

Six lenses for a diagnosis

Before discussing skills, Shengxing Yang asks the question every organisation faces: if I am happy with the current situation, why change? His thesis therefore starts with what drives change. He describes an “AI issues space”, the set of risks that automation or task augmentation (AI assisting humans without replacing them) poses to an activity, and organises these issues through six lenses.

The first, the domain lens, refers to the profession and its real problems, known only to those who practise it. The second is technical: systems are already running, and a choice must be made between buying something new and making use of past investments. The third, cognitive, assesses the company’s readiness to change, what he calls its DNA. The fourth, interactional, covers internal and external relationships, including providers’ platforms. The fifth is ethical, particularly fair use of data, debated more in Europe than in the United States or China. The sixth is managerial: transformation needs support from leadership.

To put these lenses to work, he proposes an approach called PTSD (an English acronym), which he stresses has nothing to do with a medical condition. Prioritisation [TO BE VERIFIED] is followed by trade-offs: neglecting ethics seems inconsequential at first, but can become a problem within the European framework. Then comes sensemaking, because different professions do not attach the same importance to the same function. Marketing sees a benefit; human resources sees a burden: hiring for an AI-related role can attract one hundred to two hundred CVs a day. “I am not being paid more, but I have more work. It is not fair,” he says in the recruiter’s place, who eventually refuses. Finally, dynamics: the six lenses evolve together, and one week’s priorities are not necessarily the next week’s. Diagnosis becomes a loop.

The framework offers no recipe: there is, he says, neither a right nor a wrong answer. Asked how training managers might use it, he puts awareness first, along with a question that enthusiasm tends to obscure: does the company need this transformation? AI, he believes, is a good answer, but not always the one that is needed; many companies adopt it under pressure.

What repetition teaches

Where does delegation stop? Shengxing Yang answers with examples. Industries that require the same input to always produce the same result keep large language models, which are too unpredictable, out of production. Customer service lends itself better to sharing: a conversational agent handles simple questions around the clock and passes the others to a human.

Some executives go further, he observes: no more need for junior staff, a few senior managers surrounded by agents will suffice. He accepts this for activities governed by standardised procedures, where models do better. But he defends what this reasoning erases. He says he himself learned from repetitive tasks and finds them restful after complex work. “If you push people to think about the hardest task, day after day, hour after hour, it is not possible. We are not machines.”

Yet employees seek to get rid of these tasks, even if they have to hide it. This is shadow AI, the use of tools without an employer’s approval, which he considers very difficult to control. He expresses the calculation in the first person: “From the perspective of the indicators, I have done a perfect job. I am paid thousands of euros, and I paid 20 euros.”

People still need a reason to want to learn, something he says does not come naturally: scrolling through videos provides an immediate reward, learning a tool does not. Hence the need for mechanisms within companies: promotion, pay, recognition.

Knowledge that cannot be coded

Dismissing staff to automate, then rehiring because experience is missing: when this case is raised, Shengxing Yang recalls that the scenario has already happened several times. He sees it as ethically questionable and practically risky, because some knowledge, made up of implicit conventions, is recorded in no system. “Part of culture remains embedded in the human body. It is embodied intelligence, not artificial intelligence.”

The risk extends further. Closing the door to junior staff depletes the talent pool and creates, in the medium or long term, a lock-in effect: dependence on agents that becomes difficult to escape. Hidden costs add to this, along with existing systems to integrate and a provider whose service can fail. Hence his advice for pilot projects: fail fast, iterate fast, because models evolve faster than a two-year plan.

In 2021, he recalls, research contrasted automation with augmentation. The debate now leans towards the latter, because models make recurring mistakes while expanding what a human can accomplish. Shengxing Yang shifts the question towards what the organisation knows about itself. Many executives, he observes, neglect the readability and reusability of their data. Yet what matters, in his view, is recording activity, whether codified or not: a context that an assistant such as ChatGPT cannot capture on its own.

One area remains that he would like to explore further: discretionary power, the latitude a position gives its holder, who is free to exercise their prerogatives or not, and which, he says, makes people enjoy their profession. He sees it being reallocated and warns that some people will struggle with this. “With AI, some of that latitude is delegated to tools. You then think: I am losing my power.”

Background

Shengxing Yang holds a PhD in management science. His thesis, begun in 2021 at Université Paris-Saclay, examines how organisations can prepare for AI transformation by 2035, based on a foresight exercise (the construction of future scenarios). In it, he defines an AI issues space viewed through six lenses (domain, technical, cognitive, interactional, ethical, managerial) and a diagnostic approach, PTSD, which he presented in an online publication. He holds a bachelor’s degree in administration and a master’s in international management [TO BE VERIFIED].

This article is drawn from the interview with Shengxing Yang in the podcast Human value in the age of AI (ISSEO), episode [number or title to be added]. The interview was conducted in English; quotations have been translated.