AI assistants: when should they ask for clarification in Fribourg?
In SHR — Swiss Human Resources' programme “Create your AI assistants and agents without coding”, designing the interaction deserves as much attention as generating answers. The management question is specific: in which situations should an assistant ask for clarification rather than choose an interpretation itself?
What the literature establishes
Herbert H. Clark and Deanna Wilkes-Gibbs (1986, Cognition) show, in a task involving references to figures, that establishing a mutually understood reference is collaborative work. Participants propose descriptions, adjust them and signal their understanding as the interaction unfolds. These observations suggest that an initial request should not always be treated as a complete specification.
A shared vocabulary develops through interaction
Susan E. Brennan and Herbert H. Clark (1996, Journal of Experimental Psychology: Learning, Memory, and Cognition) describe “conceptual pacts”: participants establish a shared way of referring to an object. These agreements are tied to their interaction history, not just to the general meaning of words. For an assistant, this suggests a design precaution: a familiar term does not establish that its meaning is shared.
The premature conclusion
One might conclude that an assistant should always ask a question before answering. Yet these studies establish neither the effectiveness of such a rule for AI nor an optimal number of questions. A more measured hypothesis is to reserve clarification for ambiguities that would change the expected outcome, rather than inconsequential details. (our executive and employee training programmes)
What these data cannot settle
These studies concern human interaction in experimental tasks, not generative assistants in organisations. They cannot directly establish when an interruption becomes more costly than a mistaken interpretation. However, a common practice — judging an assistant by its first answer — risks overlooking the rework required to obtain the right deliverable.
A practical check in Fribourg
In Fribourg, the programme could offer a trial using anonymised requests from the agri-food sector, bilingual industry and the university environment, some containing an ambiguous term such as “batch”, “series” or “validation”. Compare two configurations of the same assistant, one answering directly and the other allowed to ask for clarification, using comparable cases distributed among participants. Measure the proportion of deliverables requiring rework because of misinterpretation, using an assessment framework defined before the trial, together with the total time to acceptance, including clarification. Examine results separately by language and professional context: clarification would be useful if it reduced rework without exceeding the additional time deemed acceptable in advance. To go further: explore the Building AI Assistants and Agents Without Coding training in Fribourg in the canton of Fribourg, or browse our executive and employee training programmes in Switzerland.
In pictures: Building AI Assistants and Agents Without Coding in Fribourg
- building ai assistants and agents without coding
- Fribourg
- research
- training fribourg