Knowing When Not to Use AI

The rarest skill in an organisation adopting AI is not knowing how to deploy it. It is knowing when to decline

Every conversation about AI adoption assumes the relevant question is how: how to implement it well, how to govern it responsibly, how to train staff to use it effectively. A prior and more important question receives far less attention: whether to use it at all, for this particular purpose, in this particular context.

This question has largely disappeared from organisational practice, replaced by an assumption that if a capability exists and can be applied to a problem, applying it is simply good practice. That assumption is wrong often enough to deserve active resistance.

Situations That Call for Restraint

Some decisions carry consequences serious enough, and require contextual judgement subtle enough, that automating any part of them trades a small efficiency gain for a disproportionate risk. A decision about whether to remove a child from a home, whether to grant asylum, whether to end life-sustaining treatment, involves weighing factors that resist reduction to a scored input, and the appearance of algorithmic objectivity in such decisions can obscure rather than improve the quality of judgement being exercised.

Some contexts lack the data needed to support a reliable system, even when a system can technically be built. A predictive tool trained on a small, unrepresentative, or historically biased dataset will produce confident output that is not thereby made trustworthy, and the confidence of the output is precisely what makes this dangerous: a human decision-maker who is uncertain behaves differently from an algorithm that is uncertain but does not display it.

Some relationships are constituted by the encounter itself rather than by the information exchanged within it. Pastoral counsel, grief support, and the kind of mentorship that shapes a young person’s sense of their own worth depend on presence in a way that no output, however well-crafted, can substitute for. An AI system can inform a pastor’s preparation. It cannot become the encounter a person in crisis actually needs.

And some efficiency gains are not worth what they cost in trust. A customer who discovers that a personal, seemingly individualised message was entirely automated frequently feels a specific kind of betrayal that a poorly worded but genuinely personal message would not have caused, because the deception, however minor, damages the relationship more than the imperfection would have.

Why This Discipline Is Rare

Declining to use a capable tool requires an organisation to resist several converging pressures at once: the sunk cost of technology already purchased, the competitive fear of falling behind, and the genuine efficiency the tool would provide in the narrow sense of speed or cost. Each of these pressures is real, and together they make restraint feel like a failure of nerve rather than what it actually is, which is a form of discernment.

Catholic moral tradition has a name for the capacity this requires: prudence, the virtue of judging correctly what a particular situation actually calls for, rather than applying a general rule mechanically regardless of context. Prudence is not caution for its own sake. It is precision about which tool fits which situation, and the willingness to say plainly, in a specific case, that a capable and available tool is the wrong choice here.

The organisations that will be trusted longest are unlikely to be the ones that adopted AI fastest or most completely. They are more likely to be the ones that can point to specific, principled decisions not to use it, decisions that demonstrate the discernment was genuine rather than merely a limit imposed by what the technology could not yet do.