AI assistants: how should they explain a refusal in Berne?
An assistant tasked with routing internal requests soon encounters situations in which the answer is no. The management question is not simply whether that refusal is correct, but what explanation and opportunity for review should accompany it.
What the literature establishes
Jason Alan Colquitt (2001, Journal of Applied Psychology) provides evidence supporting a distinction between distributive, procedural, interpersonal and informational justice. The outcome, the procedure followed, the treatment received and the explanations provided are therefore not a single dimension of perceived fairness. Applied to an AI assistant, this distinction suggests that compliance with a rule is not enough to establish the quality of the way a request is handled.
The explanation is part of the treatment
Jerald Greenberg (1990, Journal of Applied Psychology) studied reactions to pay cuts in a field setting. The findings indicate that a thorough explanation delivered with sensitivity to those affected can mitigate reactions to an unfavourable situation without making the situation favourable. This work encourages us to view the explanation of a refusal as part of organisational treatment, rather than simply a matter of wording.
The premature conclusion
One might conclude that configuring an empathetic tone is enough to make refusals acceptable. That would confuse the quality of an explanation with its ability to secure acceptance: a courteous message corrects neither an inappropriate criterion nor a procedure that offers no opportunity for review. An assistant should communicate the applicable grounds and the actual review route, not produce a persuasive justification. (our executive and employee training programmes)
What these studies do not demonstrate
These studies concern neither language models nor agents configured without code. They do not establish that an automatically generated refusal would be perceived as just as fair as a human explanation, or that a longer explanation would be better. A common practice to avoid is therefore assessing only the politeness of the message without checking whether its grounds actually match the case.
A practical check in Berne
Within SHR’s programme « Créer ses assistants et agents IA sans coder », a relevant exercise in Berne is to configure an assistant that drafts, without making the decisions, negative responses to internal requests: access to training in the federal administration, methodological support in public health, technical intervention in telecommunications or equipment bookings in precision manufacturing. Before testing, prepare fictional cases whose applicable grounds and review options have been validated by the process owner. Then measure the proportion of responses containing three verifiable elements: grounds faithful to the case, an explicit applicable rule and a review contact actually provided for by the procedure. This indicator measures the prototype’s explanatory and procedural quality, not the fairness of the decision itself. To go further: explore the Building AI Assistants and Agents Without Coding training in Bern in the canton of 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