Most of the debate about AI in clinical development is about what the models can do. How much they can draft, how much they can review, how much they can predict? But those are conversations entirely about execution.

The more useful question, as AI moves from generating outputs to recommending decisions and acting on them, is the one that’s less spoken about. What is the human role becoming, and are we designing for it?

Judgment is becoming more valuable, not less and life sciences is already among the most advanced sectors here. In the 2026 SAS Data and AI Impact Report, with research insights provided by IDC, 84.2% of life sciences organisations reported using generative AI and 59.3% using agentic AI, level with or ahead of most other industries.

That advantage is not fixed to where the technology happens to be today. As AI systems take on more autonomous roles, the decisions they influence carry more weight, not less. The judgment applied around those decisions cannot be treated as optional at this stage and dispensable at the next. It is the constant that AI’s growing capability has to be measured against.

That is why human in the loop deserves more thought than it usually gets.

A person at the end of the process is the standard reassurance, and a reasonable one. But oversight only means something when the reviewer can see how a recommendation was reached, what assumptions sat behind it and where the model was uncertain. Given that visibility, a reviewer can do real work but without it, even a diligent one is left approving on trust alone.

The report puts numbers to the tension by showing us that almost half (48%) of life sciences organisations sit in what IDC calls the trust dilemma, where confidence in AI and its actual trustworthiness have not yet moved in step.

Being among the most mature sectors has not closed that gap. In most industries it is a commercial risk, yet in clinical development, where an output can shape a submission or a patient safety decision, getting it right matters more.

People override AI not mainly because it is wrong, but because they cannot see how it got there. The leading reasons were a lack of explanation and missing context, not inaccuracy. That reframes explainability. It is usually filed as a technical property of a model but it is better understood as the thing that lets people do their job well: trace how a conclusion was formed, question it, correct it, or approve it for a reason they can state.

That will not change as models improve. Even a system with no explainability gap left to close does not change who is accountable for what goes into a submission or a patient’s care. That responsibility was never a function of the technology’s maturity, so more mature technology does not remove it.

As AI increases the volume of outputs, human attention becomes a scarce resource. The answer is not to review everything, nor to review less. It is to direct judgment to where it changes the outcome.

Responsibility, throughout, stays human. AI can support an analysis, surface a recommendation or draft a document, but the decision remains owned by a named person, and in a regulated field accountability for what goes into evidence, a submission or a patient’s care should sit exactly there.

The organisations furthest ahead treat this as deliberate design. In the report, 91% of the life sciences trustworthy AI leaders had a responsible AI policy reaching every employee, against 7% of the laggards.

It is why workforce readiness deserves the same investment as technology. Data literacy, judgment and the confidence to question a polished-looking output are learned capabilities, and they are what the whole model rests on.

The future of clinical AI will not be determined by how much work AI can perform independently. It will be determined by how effectively organisations design systems, workflows and governance models that let people bring judgment to the decisions that matter most.