AI assistants: presenting gains or losses in Zurich?

In Zurich’s financial services, insurance and tech sectors, and in the European headquarters of international groups, an assistant can prepare a comparison without making the decision itself. The management question is specific: how can we prevent its presentation of gains and losses from quietly steering the choice?

What the literature establishes

Amos Tversky and Daniel Kahneman (1981, Science) demonstrate experimentally that different formulations of the same problem can alter expressed preferences. Presenting consequences as gains or losses is therefore not necessarily neutral, even when the options remain equivalent. For an assistant tasked with preparing a decision, preserving the data is not enough to guarantee a neutral presentation.

The reference point changes the evaluation

Daniel Kahneman and Amos Tversky (1979, Econometrica) propose a descriptive model in which outcomes are evaluated as gains or losses relative to a reference point. In this model, losses carry more weight than gains of the same magnitude. Applied to assistant design, this mechanism suggests making the chosen reference explicit: the planned budget, the current situation and the stated target are not interchangeable benchmarks.

The hasty conclusion

One might conclude that requiring positive wording, or always displaying both frames, is sufficient. Yet these studies do not establish that a single format improves every decision, or that presenting both frames eliminates bias. A more cautious rule is to make the reference point and equivalent consequences visible, without treating a different preference as an error. (our executive and employee training programmes)

What the evidence does not tell us about assistants

These studies concern human judgement, not generative assistants built without coding. They measure neither how frequently these tools change the framing nor the effects of their wording within a Zurich organisation. An internal test must also account for presentation order, familiarity with the cases and memory of an earlier answer.

A practical check in Zurich

As part of SHR’s programme « Créer ses assistants et agents IA sans coder », a Zurich team can configure an assistant to produce two versions of fictional decision cases—one focused on gains, the other on losses—tailored to financial services, insurance, tech or a European headquarters. Before testing, a designated reviewer checks that the options, probabilities and consequences are strictly equivalent and that the reference point is explicit. The same participants then review both versions in separate sessions, with the order counterbalanced across participants; the team calculates the proportion of participant–case pairs in which the chosen option changes. This rate signals sensitivity to framing, rather than proving that a decision is wrong: it helps identify presentation templates that need revision. To go further: explore the Building AI Assistants and Agents Without Coding training in Zurich in the canton of Zurich, or browse our executive and employee training programmes in Switzerland.

In pictures: Building AI Assistants and Agents Without Coding in Zurich

• building ai assistants and agents without coding
• Zurich
• research
• training zurich