A risk that is still rarely discussed
This is the most interesting study I have read on AI and its organisational consequences after deployment in a company.
It highlights a still rarely discussed risk: technological effects that vary across tasks and individuals can be transformed by managers into uniform organisational expectations.
In the experiment, AI strongly improves performance on some tasks but not on others. Yet, in the condition studied without performance feedback, only 7.8% of managers correctly anticipate this difference: 92.2% expect improvement on both tasks.
This does not mean that 92.2% of all managers will always overestimate AI. It describes a specific experiment and situation. But it reveals an essential gap between a technology whose usefulness depends on context and managerial decisions that may expect a general benefit from it.
When heterogeneous effects become homogeneous expectations
A simplified perception of beneficial AI could be summarised using three factors:
Beneficial AI at work = task × model × worker
Once AI is deployed in the real world, its benefit probably looks more like an interaction:
Beneficial AI at work = task × model × worker × organisation × deployment conditions
This is where managers’ perceptions become a prevention issue. If benefits are overestimated or generalised, the risk is that the organisation tries to ‘correct’ the worker—more training, more use, higher targets—when the difficulty may arise from the task, the model, the organisation or the deployment conditions themselves.
Adapt the organisation and the tool to real work
Better AI deployment therefore means more than adapting workers to the tool. The organisation and the use of the tool must also be adapted to real work: select the right tasks, make limitations visible, organise verification, preserve room for manoeuvre and revise objectives when expected gains do not materialise.
Occupational health can act precisely at the level of the worker, the organisation and deployment conditions: observe real activity, document differentiated effects, identify overload or loss of autonomy and help adjust the system.
Beneficial AI at work = task × model × prevention (worker × organisation × deployment conditions)
A model’s performance is not enough to guarantee the performance of work. Deployment must be assessed at the level of the task, the worker and the organisation.
Point to watch
A positive average can conceal negative effects. Assessment must distinguish tasks, workers and conditions of use before turning an average gain into uniform targets.