AI assistants: what happens to learning the job?
Building a no-code assistant makes it possible to delegate certain drafting, classification or case-preparation tasks. The managerial question is specific: how can we check that this delegation does not reduce employees’ ability to handle a subsequent case without assistance?
What the literature establishes
Norman J. Slamecka and Peter Graf (1978, Journal of Experimental Psychology: Human Learning and Memory) demonstrated, in verbal memory experiments, an advantage for items generated by participants over items they had simply read. Producing an answer and reading an answer are therefore not necessarily equivalent situations for memory. This finding does not concern AI assistants, but it cautions against confusing exposure to a good result with learning.
The paradox of a skill exercised less often
Lisanne Bainbridge (1983, Automatica) analyses a difficulty with automation: operators remain necessary for situations the system cannot handle, while their opportunities to practise may diminish. Her article offers an analysis of automated systems, not an experiment on conversational agents. The mechanism is nevertheless relevant to assistant design: removing routine cases may also remove some opportunities to learn how to recognise difficult ones.
The premature conclusion
One might conclude that all manual work must be retained to preserve professional expertise. These studies do not justify such a rule, nor do they demonstrate that using an assistant necessarily causes a loss of skill. A more specific design hypothesis is to retain a stage in which employees formulate their own diagnosis before consulting the assistant’s proposal. (our executive and employee training programmes)
What the evidence cannot establish
Verbal memory experiments and an analysis of industrial automation cannot quantify the effects of a generative assistant on professional learning. In practice, measuring only the quality of the completed case also leaves a question unanswered: does that quality belong to the employee, the system or their combination? An unassisted assessment using a new, comparable case allows a different dimension to be examined beyond assisted performance.
A practical check in Berne
As part of SHR’s programme « Créer ses assistants et agents IA sans coder », a pilot in Berne could use fictional or authorised case materials relevant to the federal administration, public health, telecoms or precision manufacturing. Comparing two randomly assigned conditions — an immediately visible proposal or a personal diagnosis written before the proposal is displayed — would allow this design choice to be tested. Before and after the pilot, participants would handle new cases of comparable difficulty without an assistant, with the proportion of correctly justified decisions assessed against a predefined professional rubric. Changes in this proportion under each condition, scored without knowledge of group allocation, would provide a local indicator of learning rather than general proof about AI. To go further: explore the Building AI Assistants and Agents Without Coding training in Bern, or browse our executive and employee training programmes in Switzerland.
In pictures: Building AI Assistants and Agents Without Coding in Bern
- building ai assistants and agents without coding
- Bern
- research
- training bern