When people talk about preparing doctors for the AI era, the conversation often goes straight to tools: which model, which vendor, which use case. Those questions matter. But there is an earlier question that may be more useful: what kind of problem is this system actually presenting?
Medicine is full of problems that look similar from a distance and behave very differently up close. Treating each one as a technical problem is an easy way to create a technically impressive solution that does not improve much.
Not every problem needs the same kind of intelligence
A useful systems-thinking shorthand separates situations into clear, complicated, and complex domains. The labels are less important than the discipline of asking which one we are in.
Clear work has stable rules and an observable answer. Checking whether required images arrived, routing a routine study, or flagging a missing protocol element can often be standardized or automated—provided someone monitors the exceptions.
Complicated work still has an answer, but it may require expertise, additional information, or tradeoffs. Technical image quality, protocol selection, and many imaging workups fit here. AI can summarize, organize, or make patterns easier to see. The radiologist still has to decide what matters and what to do next.
Complex work is different. A finding may be clear, while its implications are not. Comorbidities, access to care, patient preferences, changing evidence, communication, and local capacity interact. Cause and effect may only become clear later. In these situations, a confident recommendation can be useful; it is not the same thing as a reliable resolution.
AI readiness is a systems skill
This is where “AI literacy” needs to become more than prompt-writing or knowing the vocabulary. A tool entering a clinical workflow changes the workflow. It redistributes attention, creates new handoffs, and may move an error somewhere less visible. A model can perform well on a validation set and still be poorly matched to the people, incentives, data quality, or follow-up pathways around it.
So the practical question is not simply, “Can this be automated?” It is also:
- What decision is the tool influencing?
- Who sees, verifies, and acts on its output?
- What happens when the data, workflow, or prevalence changes?
- What outcome would tell us that care is actually better?
Those questions are not a reason to avoid AI. They are how we keep a pilot from becoming a new source of friction or risk. In a clear domain, a reliable process and exception monitoring may be enough. In a complicated domain, expert review and local validation matter. In a complex domain, the sensible approach is usually a small, reversible test with feedback from the people doing the work.
Small experiments, real feedback
That last point may be the most important. Complex systems are not “solved” once. They are navigated: try something modest, observe what changes, listen to the people affected, and adjust. This is familiar quality-improvement work, even if the current vocabulary is AI governance.
The regulatory and educational landscape is moving in the same direction. The FDA’s 2025 draft guidance on AI-enabled device software frames evaluation across the product lifecycle, including post-market performance. The AAMC’s AI competencies initiative similarly treats responsible use as a professional and team-based capability, not merely a technical one.
For radiologists, perhaps the goal is not to become data scientists or to treat every difficult problem as an AI opportunity. It is to recognize when a tool fits the problem, when it needs guardrails, and when the most valuable intervention remains a better conversation among the people in the system. That is systems thinking. It may also be one of the more durable ways to prepare for whatever comes next.