When an AI prescription becomes a real intervention

The transition from a prediction to a prescription changes the question a reader should ask. The output is no longer only a number to compare with a later observation; it helps determine what happens next. That makes the decision pathway itself part of the evidence.

An intervention inside a clinical protocol

Chen and colleagues reported a prospective daGOAT study in September 2025. In the analysed group of 110 haploidentical-transplant recipients, the system identified 57 for additional ruxolitinib, with 56 starting immediately. Its conditional automation allowed physician intervention. The authors explicitly state that the study was not designed to establish the regimen's efficacy or superiority over physicians' intervention decisions.¹

This is human clinical research, not a simulated prescribing benchmark. Its principal lesson for this article concerns implementation and acceptance, with treatment-effect claims kept separate.¹

A prediction changes its own future

Our interpretation starts with a simple problem. Suppose a risk forecast prompts an intervention, and the feared event does not occur. Was the original forecast wrong, or did the intervention prevent the event? The observed outcome alone cannot settle that question. A system that acts on its predictions needs an evaluation designed around the action, not only the forecast.

The same reasoning applies when comparing treated and untreated people. If the groups were selected because their predicted risks differed, their later outcomes cannot be read as though treatment had been allocated independently of those risks. We would ask how the comparison deals with the reasons people entered each group, and which unmeasured differences might remain.

Make the human boundary visible

For an AI prescribing system, we would want the operational record to identify the exact responsibility delegated. Did the system decide whom to consider, whether to begin an intervention, or how to adapt it? We would record the moments when a clinician changed the plan and the reasons for doing so. Such decisions are part of the clinical process, not noise to remove from the narrative.

In an imagined follow-up evaluation, we would predefine how disagreements are assessed. A disagreement might expose missing information, a legitimate clinical exception or an unreliable prediction. It should trigger investigation rather than an assumption that either the algorithm or the clinician must be right.

The evidence worth asking for next

Our preferred next report would connect implementation measures with patient outcomes under a design that supports the intended comparison. It would specify the population to which the conclusion applies and preserve the cases in which the workflow stalled or was overridden.

The useful achievement here is that a bounded AI-mediated intervention was examined in real care. The larger claim—that using such a system improves treatment—requires its own evidence. Keeping those conclusions distinct gives readers a way to appreciate the work while understanding the question it leaves open. Nothing in this report is a recommendation for an individual patient's treatment.