Essay

Michel de Certeau’s tactics reveal why frontier LLMs are never neutral tools: every workaround users invent inside a platform’s ‘proper place’ can be captured, generalized, or closed off.

AI can loosen authorship without loosening accountability: the person who signs an output may lose credit for producing it yet become the system’s moral crumple zone when it fails.

Failed rollouts do more than create scepticism: they teach employees that action is futile, turning the next AI initiative — however sound — into a test of trust already depleted.

Every agent roadmap begins with work it cannot see: people translate tacit practice into legible inputs, then absorb the exceptions, repairs, and coordination that keep automation running.

Automation can make a workforce look more capable while preventing competence from forming: people inherit outputs without the judgment, calibrated trust, or ability to take back control.

Chat narrowed the skills gap by supplying answers. Agents widen it again: they reward those who can turn ambiguous work into briefs, constraints, and evaluable sequences.

The AI super-user is less a technical type than the visible beneficiary of Bourdieu’s linguistic capital: LLMs reward the ability to formulate, frame, and judge — then turn class advantage into performance.

AI does not simply extend thought; it changes the plane on which thought moves, opening fast, rhizomatic exploration while weakening the intuitive ground that lets us know where we are.

Agentic AI removes the machine bottleneck only to expose another: the finite human capacity to set direction, retain context, verify outputs, and decide what deserves trust.

GenAI adoption is not a technology problem with a training solution. It is a social adoption problem, carried by the people between leadership and frontline peers whose trust and professional identity determine whether use sticks.

Feedback is not a free good: more than a third of interventions lower performance, making the design of the system — not good intentions — the real management problem.

Underperformance triggers control; control makes honest information costly; the organisation then loses the very signals it needs to understand the decline.