IA on the job: learning from projects without forgetting failures
Organisations want to use AI to turn their project archives into reusable recommendations. The management question is specific: how can an assistant trained on success-heavy records be prevented from turning insufficiently tested practices into internal recipes?
What the literature establishes
Jerker Denrell (2003, Organization Science) uses theoretical analysis to show how the underrepresentation of failures can distort organisational learning. When organisations mainly observe survivors, certain practices may appear conducive to performance without that observation establishing their effectiveness. Applied to a document-based assistant, this reasoning suggests examining how records were selected before assessing the quality of its recommendations.
The mechanism that sustains selection
James G. March (1991, Organization Science) analyses the tension between exploiting existing knowledge and exploring new possibilities. His model shows how the more immediate benefits of exploitation can undermine long-term adaptation. An assistant that systematically recirculates methods from projects recognised as successes could reinforce this dynamic; this is a proposed application, not an established finding about generative AI.
The premature conclusion
One might conclude that simply adding abandoned projects to the document repository would solve the problem. Yet abandonment may result from a portfolio change, a regulatory constraint or a commercial decision, without invalidating the method used. The circumstances and reasons behind the outcome therefore need to be documented, rather than treating the categories of success and failure as explanations. (our executive and employee training programmes)
What the available evidence cannot establish
These studies measure neither the quality of a current assistant nor the causal effect of a more diverse corpus on decisions. In practice, discontinued projects may also leave less complete records than completed projects, making comparisons difficult. An accurately referenced answer can therefore still rest on a selective record of organisational experience.
A practical check in Bâle
As part of IA on the job, a Basel-based pharmaceutical or life sciences team could select a group of projects connecting international R&D with Rhine logistics, then compare the project register with the corpus actually accessible to the assistant. The measure would be documentary coverage by outcome: the number of projects with a usable review in the corpus, divided by the number of registered projects in each category — completed, discontinued or redirected. The register and the criteria for a usable review would be fixed before counting, within a scope that respects access permissions. A coverage gap would not demonstrate bias in every answer, but it would give management a specific shortcoming to address before presenting recommendations as lessons from collective experience. To go further: explore the AI on the Job training in Basel in the canton of Basel-Stadt, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Basel
- ai on the job
- Basel
- research
- training basel