IA on the job: what record should an AI-assisted decision leave?
A commercial proposal or an internal memo can now be prepared with AI and approved without preserving the reasons behind the choice. The specific management question is this: what minimum record should be required to keep an assisted decision intelligible when someone else takes over the file?
What the literature establishes
James P. Walsh and Gerardo Rivera Ungson (1991, Academy of Management Review) propose a theoretical framework for organizational memory: information from the past is retained in several repositories, including people, routines and structures. Their analysis distinguishes the acquisition, retention and retrieval of information, rather than treating memory simply as an archive. Applied to AI, this framework invites a distinction between saving an answer and being able to retrieve the reasons for a decision.
Having to explain does not always improve reasoning
Jennifer S. Lerner and Philip E. Tetlock (1999, Psychological Bulletin) show in a literature review that the effects of accountability depend, among other things, on its timing and the anticipated audience. Depending on the conditions, it can encourage closer examination or the justification of an already adopted position. Asking for decision criteria before approval is therefore not equivalent to requesting an explanation once the outcome is known.
The premature conclusion
One might conclude that every exchange with AI should be archived. These studies establish neither the necessity nor the effectiveness of such an arrangement. A short record linking the choice, the information checked and the person approving it is instead a working hypothesis to test, not a proven recipe. (our executive and employee training programmes)
What the evidence cannot guarantee
Both publications predate generative AI and do not measure the effect of a decision record in a company using it. An explanation written afterwards can be coherent without reflecting the reasoning actually followed. Moreover, keeping more text can increase the amount of sensitive data stored without making a file easier to take over.
A practical check in Lugano
As part of “IA on the job”, a team in Lugano could test an Italian-language record for two weeks for one type of non-sensitive decision: the chosen option, criteria, information checked and person who approved it. The case would be selected to suit the business — banking, fashion, trading or an Italian-speaking family SME — while respecting internal confidentiality rules. At handover, a colleague not involved should be able to retrieve all four elements without contacting the author; the proportion of tested records that allow this would provide a verifiable measure of their documentary usefulness, not of the correctness of the decisions. To go further: explore the AI on the Job training in Lugano, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Lugano
- ai on the job
- Lugano
- research
- training lugano