Physicians are adopting AI faster than anyone can explain who is accountable for it. Doximity's 2026 State of AI in Medicine report found 94 percent of physicians are already using AI or interested in it, but only 8 percent say their organization's AI decision-making process is clear, and 47 percent say it is still evolving. A similar adoption-versus-readiness gap shows up at the system level too: HIMSS and Guidehouse's 2026 Healthcare AI Trends report found 78 percent of health systems are engaged in AI projects, but only 52 percent feel operationally ready. The gap is not whether AI belongs in clinical care. It is whether a health system can show, all at once, where a clinician's judgment must override the system, whether that override is ever reviewed, and who is accountable for it.
Clinicians did not stop trusting AI. Most of them never got a clear answer about who is accountable for it.
What physicians themselves say
Doximity's 2026 State of AI in Medicine report, surveying 3,151 US physicians across 15 specialties, found that 94 percent are already using AI or interested in doing so. At the same time, only 8 percent said their organization's AI decision-making process is clear, and 47 percent said it is still evolving. The top concern physicians raised was not the technology itself. It was accuracy and reliability, cited by 71 percent, the exact question a documented override protocol is built to answer.
Is system-level readiness catching up?
Not much faster, though for a wider set of reasons. HIMSS and Guidehouse's 2026 Healthcare AI Trends report, based on a survey of 50 healthcare leaders published in February 2026, found that 78 percent of health systems are engaged in AI projects, while only 52 percent feel operationally ready to implement them at scale. The report attributes that gap to inconsistent data quality, governance, cybersecurity risk, and staff alignment together, not one single cause. Governance is part of the readiness gap, not the whole of it, but it is the part a health system can close fastest and most directly.
What's being built to close the gap
Real infrastructure is emerging. The Coalition for Health AI (CHAI), built through collaboration across more than 100 healthcare organizations, including academic medical centers, regional care facilities, and community health centers, has released governance playbooks and a Risk Categorization Tool for AI oversight. That work matters. It also has not yet closed the gap the Doximity and HIMSS data show, which means the responsibility sits with each organization to put that infrastructure into practice, not wait for it to arrive fully built.
The G.U.A.R.D.™ test of an AI-governed health system
Being an AI-adopting health system means the tools are in use. Being an AI-governed one means leadership can prove five things, drawn from the G.U.A.R.D. Framework™:
1. Governance. Is there a named person with documented, written authority to pause an AI system the moment it raises a patient safety concern? Neither ISO/IEC 42001 nor the NIST AI Risk Management Framework mandates that exactly one person hold this authority, but a named owner is the clearest way to demonstrate what both are actually asking for: documented, unambiguous accountability. 2. Understanding. Can the organization produce a current inventory of every AI tool active in clinical, diagnostic, and documentation workflows, including the ones clinicians adopted on personal devices? 3. Authority. Has the Human Authority Line™ been drawn for each clinical workflow: the point where an AI recommendation ends and a clinician's judgment must remain non-delegable? 4. Reputation. When a clinician overrides an AI recommendation, is that override captured and reviewed, or does it disappear the moment the clinician moves on? 5. Design. Once built, does your own clinical and compliance leadership fully own and operate this architecture?
Falkovia's point of view
An AI recommendation that appears inside the EHR, next to real lab results, inherits the trust clinicians already place in that workflow before anyone has established that the recommendation deserves it. I call this trust transfer, and it is exactly how a 94 percent adoption rate and an 8 percent clarity rate end up sitting in the same report. The tool earned a place in the workflow. Accountability for what it recommends did not automatically come with it.
The stakes are also changing shape. As AI systems move from advising to acting inside clinical workflows, the relevant question is no longer only whether a recommendation was right. It is whether a named person had the authority, in writing, to stop the action before it happened, and whether the record of that override gets reviewed rather than disappearing the moment a clinician moves past it. A health system does not need to wait for a lab-grade safety net to build both pieces at once.
That architecture, a documented line, a reviewed override record, and someone who answers for both, does not require waiting for CHAI's playbooks to mature or for state legislation to settle. It is buildable now, inside the workflow, before the next AI-assisted decision is the one a board or a regulator asks about.