IA on the job: when should a pilot be stopped?
After choosing an assistant and adapting a few processes, stopping the trial can feel like wasting the work already done. The management question is more precise: how can a decision to continue an AI pilot avoid treating past effort as an argument in its favour?
What the literature establishes
Hal Arkes and Catherine Blumer (1985, Organizational Behavior and Human Decision Processes) documented the sunk-cost effect: a prior investment of money, time or effort can increase the tendency to continue a course of action. A desire not to appear wasteful helps explain this effect. For an AI pilot, hours already spent on configuration are therefore not, by themselves, a reason to continue.
When initial responsibility shapes the next decision
Barry Staw (1976, Organizational Behavior and Human Performance) showed experimentally that personal responsibility for an initial decision could encourage further investment following negative results. Continuing may then serve to justify the earlier choice, rather than respond solely to future prospects. Applied to AI, this mechanism suggests that a pilot’s sponsor should not be its only reviewer.
The premature conclusion
These studies do not establish that a disappointing pilot should be stopped immediately. An initial difficulty may still lead to useful learning, and switching tools may entail genuine future costs. The relevant distinction is between past effort that cannot be recovered and expected benefits and expenditure that can still be avoided. (our executive and employee training programmes)
What the evidence cannot settle
These classic studies do not concern generative AI and provide no universal stopping threshold. They illuminate a risk in reasoning, without showing that all persistence amounts to escalating commitment. In practice, a review that totals the hours invested without specifying conditions for continuation makes this risk difficult to examine.
A practical check in Lugano
Within SHR’s “IA on the job” programme, one exercise is to set a continuation condition and a review date before the trial: for example, continue only if total preparation and correction time is lower than for the usual process, with no increase in predefined instances of non-compliance. In Lugano, this comparison could cover Italian-language drafts of banking correspondence, fashion product descriptions, requests for quotations in trading or customer replies in an Italian-speaking family business, using comparable cases in an authorised environment. At the review, someone who did not sponsor the initial choice checks time records and instances of non-compliance, then records whether the condition has been met. Any extension despite a failure to meet it must specify a testable future benefit and a new deadline, rather than invoke work already completed. To go further: explore the AI on the Job training in Lugano, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Lugano
- ai on the job
- Lugano
- research
- training lugano