AI pilots: when should a manager stop?

Launching an artificial intelligence pilot is a visible decision; ending it is equally visible, especially for the manager who championed it. The useful question is specific: how can a manager decide whether to continue without turning past expenditure into a justification for new spending?

What the literature establishes

Barry M. Staw (1976, Organizational Behavior and Human Performance) showed experimentally that responsibility for an initial decision could encourage further financial commitment following unfavourable results. This finding describes a mechanism of escalating commitment, not a general inability among managers to acknowledge mistakes. For an AI pilot, it invites scrutiny of whether continuation rests on future prospects or on the need to defend the launch.

The weight of what cannot be recovered

Hal R. 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. Their research highlights, among other factors, the desire not to appear wasteful of resources already committed. Applied to AI, this mechanism suggests that tool configuration or training hours can become arguments for retention, even when they say nothing about future benefits.

The premature conclusion

One might conclude that a disappointing pilot should be stopped immediately. That would confuse an irrecoverable expense with useful information: a trial may reveal a remediable condition of use or generate learning relevant to the next step. The distinction concerns the justification for that step: expected benefits and costs, rather than a determination to make past expenditure pay off at any price. (our executive and employee training programmes)

What these findings cannot decide

These studies concern neither generative AI nor deployments in organisations in Valais. They provide no universal threshold for stopping a project, nor do they establish that continuing after a setback is necessarily irrational. A managerial assessment must also consider future exit costs, alternatives and the value of a clearly bounded additional trial.

A practical check in Sion

Within SHR — Swiss Human Resources’ programme « Manager à l'ère de l'intelligence artificielle », one exercise is to prepare a brief before the trial, specifying the expected outcome, success threshold, review date and stopping conditions. In Sion, this could concern a non-critical documentation task in hydropower, administrative support without patient data in healthcare, a wine product sheet or visitor information for Alpine tourism. On the agreed date, measure the proportion of outputs usable without substantial correction against a rubric defined at the outset, then compare the result with the stated threshold and record the costs still to be incurred. Someone who did not champion the launch reviews the decision to continue, modify or stop: the check concerns consistency between the original criteria, observed results and written justification, not the amount already spent. To go further: explore the Managing in the Age of Artificial Intelligence training in Sion in the canton of Valais, or browse our executive and employee training programmes in Switzerland.

In pictures: Managing in the Age of Artificial Intelligence in Sion

• managing in the age of artificial intelligence
• Sion
• research
• training sion