Clinician trust is still the gatekeeper for healthcare AI
Healthcare organizations are no longer asking whether artificial intelligence belongs in clinical and administrative workflows. They are asking which tools deserve scarce capital, which can be trusted at the bedside, and which will create more work than they remove. At a Healthcare Dive virtual event, industry experts drew a blunt picture of adoption: enthusiasm is real, but implementation still depends on governance, costs, and clinician confidence.
The adoption problem starts with the people expected to use the software. In a 2023 American Medical Association survey of 1,000 physicians, 41% said they were equally excited and concerned about healthcare AI tools. That split reflects a central hazard in clinical informatics: a tool that is technically capable but socially distrusted will stall long before it reaches scale. When clinicians worry that a model may miss nuance, misread context, or produce a recommendation that does not fit a specialty workflow, the default response is hesitation. In high-acuity settings, hesitation can function as a safety check.
A practical path to buy-in emerged from Mayo Clinic’s experience. Micki Tripathi, the organization’s chief artificial intelligence implementation officer, said tools gain more traction when clinicians help build them. A nurse-to-nurse handoff summary tool achieved strong adoption because nurses themselves identified the need, conceptualized the product, and helped roll it out among colleagues. In medicine, that sequence matters because adoption often follows authorship. Staff are more likely to use a tool they helped shape than one handed down from a distant technology team.
Specialty medicine makes the challenge sharper. Tripathi pointed to cardiology as an example, where even a strong general-purpose model may not be specialized enough to match the complexity of a department with 43 cardiac divisions. Clinical expertise is not interchangeable across subspecialties, and AI tools that flatten those distinctions can lose legitimacy fast. A cardiology workflow that spans heart failure, electrophysiology, structural intervention, and imaging cannot be served well by a generic interface that misses subspecialty context.
One sentence captures the adoption problem plainly: clinicians will not trust a tool that ignores how they actually practice.
Rural hospitals are being pushed toward administrative AI first
Cash-strapped hospitals face a different calculus. They may need AI most, yet have the least room to absorb the upfront cost of implementation, training, and maintenance. For rural systems in particular, the value proposition has to be immediate and operational rather than aspirational.
Dr. Alexander Fortenko, vice president of innovation at Ochsner Health, said hospitals with limited budgets should start with administrative technologies, especially tools that manage prior authorizations. That recommendation reflects the pressure points in a strained delivery system. Prior authorization consumes staff time, slows treatment, and creates friction between patients and providers. If AI can reduce manual bottlenecks in that process, the return appears on both sides of the encounter: fewer delays for patients and less clerical drag for clinicians.
Lydia Daniels, vice president of analytics and AI at Epic, argued that revenue cycle management may offer an even clearer economic case. Unlike some clinical AI products whose benefits are hard to quantify, revenue cycle tools can translate into direct operational gains. Coders can handle more volume and more specialties without necessarily shrinking the workforce. In rural medicine, Daniels said, doing more with the same staff is not a slogan; it is the operating model.
The financial logic is straightforward, but the clinical consequence is equally important. Hospitals with thin margins cannot afford AI projects that create complexity before they create relief. Tools that reduce administrative burden are more likely to survive procurement review because they connect to a measurable pain point. A hospital may tolerate a slower path to clinical decision support. It is far less likely to tolerate software that adds another layer to an already fragile billing and authorization process.
AI savings remain uncertain while payers and providers press their cases
The industry still lacks a clean answer to a question that sits underneath nearly every AI purchasing discussion: does healthcare AI actually lower costs?
The answer is not settled. Some payers argue that provider-side AI may increase spending by helping organizations document and code more conditions during patient evaluation. In that view, the technology does not merely uncover missing complexity; it can also intensify the billing arms race between insurers and health systems. The result is a system in which AI becomes both efficiency tool and financial weapon, depending on who is using it.
Shri Santhanam, chief product officer at Machinify, framed the dispute as part of a longer-running conflict over claims and reimbursement. The adversarial relationship between payers and providers predates AI by decades, but new tools sharpen the contest because they can automate analysis at scale. A model that speeds claims review on one side invites a countermeasure on the other. Each iteration can increase throughput while also hardening suspicion.
That dynamic has consequences beyond contract negotiation. When organizations adopt AI primarily to gain leverage over another stakeholder, clinical and administrative value can blur. A system optimized to win reimbursement battles may not necessarily improve patient care. The more AI gets embedded in payment workflows, the more the industry will need evidence showing where efficiency ends and cost shifting begins.
Federal inaction is forcing hospitals into a state-by-state compliance maze
Governance may become the most decisive factor of all. In the absence of a comprehensive federal framework, healthcare organizations are confronting a patchwork of state laws that differ in scope, speed, and specificity. David Arnold, senior vice president and chief AI and data officer at Hackensack Meridian Health, said there is no single federal regulatory structure in place, while state-level activity is moving quickly.
For multistate systems, that creates an operational burden. A health system operating across state lines cannot easily maintain one uniform AI policy if local requirements differ. Aashima Gupta, global director of healthcare at Google Cloud, said that kind of fragmentation can make governance unsustainable for larger organizations. In practice, compliance teams may need separate review pathways, legal checks, and documentation standards depending on geography.
Industry groups are trying to fill the void. Lucy Orr-Ewing, director of policy and strategy at the Coalition for Health AI, said the most useful early investment is education. CHAI focuses on AI literacy for nurses, physicians, cybersecurity staff, IT teams, and data personnel so they can identify errors early and understand where models may fail. That multidisciplinary model recognizes a basic truth of healthcare deployment: safe AI is not just a software issue. It is a workforce issue, a data issue, and a governance issue at the same time.
The next phase of adoption will not be decided by enthusiasm alone. Hospitals will need tools that clinicians can trust, budgets that can support implementation, and governance structures strong enough to survive a fractured regulatory environment. Without those pieces, even useful AI will remain trapped in pilot programs and procurement decks.