UC San Diego Health offers a local lens for a larger question: what safeguards do physicians need before AI leaves the pilot stage and enters routine clinical workflow?

Health systems are no longer asking only whether artificial intelligence can work in medicine. They are asking how AI should work once it touches daily care.

UC San Diego Health offers a local lens on that question, but the public sources reviewed do not yet confirm which AI tools the system has moved from pilot testing to routine clinical operations. That distinction matters. A pilot can test a tool in a narrow setting. An operating model must answer harder questions: who approves the tool, who monitors it, who reviews its output, who corrects mistakes, and who remains accountable when AI shapes the clinical record or patient communication.

For physicians, the central issue is not the algorithm. It is the workflow around the algorithm.

Across health systems, AI tools may support tasks such as drafting notes, summarizing conversations, prioritizing messages, flagging risks, assisting with coding, scheduling care, or supporting population health outreach. Ambient documentation tools, for example, use clinician-patient conversations to generate draft clinical notes for clinician review. But draft support does not remove physician responsibility. Clinicians still need clear rules for review, correction, signature, documentation standards, patient consent, privacy, and liability.

That is where pilots become real operating questions. If an AI-generated note contains an error, who catches it? If a patient message uses the wrong tone or omits a warning sign, who is responsible for that failure? If a tool works well in one specialty but poorly in another, who monitors that difference? If physicians spend less time writing but more time checking AI output, has the system reduced the burden or just shifted it?

Those questions should shape any health system’s AI transition plan. Physicians need to know whether AI tools integrate with the EHR, whether the system stores audio or transcripts, whether patients can opt out, whether compliance teams audit performance, and whether leaders track metrics such as note-closure time, inbox burden, documentation accuracy, safety events, clinician satisfaction, and patient complaints.

Academic health systems face added complexity. They often research pilots, quality-improvement projects, vendor deployments, and clinical operations side by side. Physicians need clear lines between experimentation and routine care.

The safest conclusion for now is also the most useful one: before any health system treats AI as everyday infrastructure, physicians need a transparent operating model. That model should define procedures for approval, consent, review, monitoring, error correction, data use, accountability, and shutdown.