Pharmaceutical investigations are difficult because the obvious explanation is not always the correct one.

An event occurs. Production needs an answer. Quality needs an investigation. Timelines start running. Under pressure, there is a natural temptation to settle on a plausible explanation and move towards closure.

This is precisely where AI could become useful.

AI can analyse information at a scale humans cannot easily replicate. It can compare an event with hundreds of historical deviations, identify similarities, retrieve relevant maintenance or process information and highlight patterns across products, equipment, shifts or locations.

That could significantly improve the breadth of investigation.

Imagine an investigator examining an intermittent equipment failure. AI identifies six apparently unrelated deviations over the previous eighteen months involving the same subsystem, three maintenance interventions and a subtle change in process behaviour preceding each event.

That does not identify the root cause.

But it gives the investigator somewhere valuable to look.

AI can also act as a challenge mechanism. If an investigation concludes “operator error”, an AI-supported review could ask whether similar errors have occurred across different operators, whether the procedure is difficult to follow, whether workload or interface design contributed, and whether previous retraining actions actually prevented recurrence.

Used properly, AI could make investigations less dependent on individual memory and more connected to organisational knowledge.

But there is a danger.

AI produces plausible explanations extremely well.

A plausible explanation is not evidence.

Pharmaceutical investigations therefore need a clear boundary between hypothesis generation and causal determination. AI may be excellent at the first. Human scientific judgement, supported by evidence, should remain central to the second.

Current regulatory thinking is consistent with this principle. FDA’s draft framework for AI supporting drug regulatory decision-making is based on assessing credibility for a defined context of use rather than treating AI output as inherently reliable. EMA likewise highlights risks around bias, reliability and trustworthiness when AI is used across the medicinal-product lifecycle.

Perhaps, therefore, the best role for AI in investigations is not to provide the answer.

It is to help investigators ask better questions.

And better questions are often where better root cause analysis begins.