AI on the job: does a well-written procedure describe the work as it is done?
Formalising an activity with an assistant can quickly create the impression that its execution is under control. The managerial question is more precise: what needs to be checked before approving a procedure drafted with AI?
What the literature establishes
John Seely Brown and Paul Duguid (1991, Organization Science) draw on research into work practices to show that formal descriptions of work can leave out essential elements of how it is accomplished. Exchanges between colleagues and adjustments made in context are not simply deviations to eliminate: they can make the activity possible. For a procedure drafted with AI, this suggests checking what the document omits, not only what it states.
A routine is more than its description
Martha S. Feldman and Brian T. Pentland (2003, Administrative Science Quarterly) distinguish the abstract dimension of a routine from its concrete enactments by people in particular situations. Their theoretical framework explains how a routine can generate both stability and change. A procedure can therefore provide guidance without capturing every condition needed for its execution.
The premature conclusion
One might conclude that procedures are useless or that local practices should always take precedence. These studies justify neither abandoning formalisation nor legitimising every deviation. The task is to distinguish a requirement that must be met, a missing condition that needs to be made explicit, and a practice that needs correcting. (our executive and employee training programmes)
What these studies do not establish
These articles do not address generative AI or measure the reliability of procedures it has produced. They provide an analytical framework, not evidence that AI-assisted drafting is inferior to human drafting. A document review alone nevertheless remains insufficient to establish whether a text covers actual execution conditions.
A practical check in Neuchâtel
As part of SHR’s “IA on the job” programme, a Neuchâtel team working in microtechnology, watchmaking or precision microelectronics could test this point on a draft dimensional inspection instruction, without changing the approved procedure. During an authorised review at the workstation, an operator and a quality manager would record every step requiring information, a condition or a judgement absent from the draft, then classify the gap rather than automatically endorsing the observed practice. The measure would be the proportion of reviewed steps with at least one gap, each supported by a recorded observation; repeating the review after revision would check whether the identified omissions had been addressed, without claiming to certify the entire procedure. To go further: explore the AI on the Job training in Neuchâtel, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Neuchâtel
- ai on the job
- Neuchâtel
- research
- training neuchâtel