AI on the job: simulating a customer is not the same as consulting one

In Zurich’s financial services, insurance, tech companies and European headquarters of international groups, an assistant can help anticipate reactions to an offer or a service change. The management question is specific: under what conditions can these simulated reactions inform a decision without taking the place of consultation?

What the literature establishes

Nicholas Epley, Boaz Keysar, Leaf Van Boven and Thomas Gilovich (2004, Journal of Personality and Social Psychology) show that understanding another person’s perspective can involve anchoring on one’s own perspective and then adjusting insufficiently. Putting oneself in a customer’s shoes therefore does not guarantee enough distance from one’s own standpoint. For AI use, these findings encourage scrutiny of the assumptions built into the prompt, without demonstrating that the model reproduces this human mechanism.

Getting a perspective rather than imagining it

Tal Eyal, Mary Steffel and Nicholas Epley (2018, Journal of Personality and Social Psychology) distinguish the effort to imagine another person’s perspective from gaining access to it through conversation. In their experiments, mentally putting oneself in another person’s position does not consistently improve judgment accuracy, whereas directly eliciting their perspective can improve it. The useful distinction is between generating a plausible hypothesis and obtaining information from the person concerned.

The premature conclusion

It would be excessive to conclude that simulating a customer with AI is useless. A simulation can help prepare questions or consider objections; this is preparation, not validation of market expectations. Within SHR’s “IA on the job” programme, the learning objective can therefore be framed as turning a simulated reaction into a question to investigate, rather than an established fact. (our executive and employee training programmes)

What the evidence does not establish

These studies concern human judgment, not generative models’ ability to represent Zurich customers. They neither quantify the reliability of a synthetic persona nor establish that an interview alone reveals future behaviour. One practice nevertheless remains questionable: treating several simulated responses as so many independent accounts from real customers.

A practical check in Zurich

For a financial offering, an insurance customer journey, a tech service or a project led by a European headquarters in Zurich, record before the interviews the expectations AI attributes to customers and the decisions they might influence. Then elicit relevant customers’ priorities in their working language, using open questions before showing them the simulated hypotheses. Measure the proportion of predicted expectations mentioned spontaneously, recording contradictions and priorities missing from the simulation separately. This check can serve as an “IA on the job” exercise: it measures agreement within the sample consulted, not representativeness across the market. To go further: explore the AI on the Job training in Zurich, or browse our executive and employee training programmes in Switzerland.

In pictures: AI on the Job in Zurich

• ai on the job
• Zurich
• research
• training zurich