By Leigh Burchell (Altera Digital Health), Chair, EHR Association

At a recent EHR Association General Membership Meeting, an event we hold monthly to bring together our members for conversations on a wide variety of topics, I had the pleasure of sitting down with Bret Borota, Head of Global Strategic Sales and Business Development for Healthcare and Payer Segments at Amazon Web Services (AWS), for a wide-ranging fireside chat on the state of AI in health IT.

Bret brings a unique vantage point to this conversation: a current portfolio spanning medical imaging, EHRs, interoperability, revenue cycle management, and commercial payers, plus 20 years in health IT before joining AWS.

Where AI Is Working Today in Health Care

I opened by asking Bret where he’s seeing the earliest, most meaningful wins. His answer was straightforward: success today is still largely persona- and department-specific, rather than sweeping.

In Bret’s opinion, these areas stand out as showing the most early traction:

• Ambient and clinical documentation, where generative AI has driven strong, sticky adoption.
• Medical imaging, arguably the fastest-growing use case in both diagnostics and radiologist workflow prioritization.
• Revenue cycle management, where claims automation and coding tools are being embraced on both the provider and payer sides.
• Enhanced patient engagement in a variety of ways, easing their care experience.

The common thread across all three was that AI is most successful when it fits within an existing, familiar workflow.

The Next Horizon: Agentic AI

Bret also pointed us toward a more transformative future built around agentic AI (AI that takes purposeful actions on its own to achieve goals) systems that communicate across departments, specialties, and platforms to suggest and execute “next best actions” that transcend any single workflow. The opportunity isn’t just augmenting today’s processes; it’s reimagining what workflows could look like once AI is no longer constrained by legacy system boundaries.

Where to Focus Investment

We also talked through how software developers should optimize their AI investments given how quickly both the technology and the competitive landscape are shifting. Bret’s advice centered on building closed-loop partnerships between the technology developer, data source, and end user.

Without direct end-user input and real-world data feedback embedded in the development cycle, he cautioned, AI models will fail to find strong adoption or sustained use.

Without direct end-user input and real-world data feedback embedded in the development cycle, he cautioned, AI models will fail to find strong adoption or sustained use.

Governance First: Bring the Model to the Data

One of the most practical takeaways from our conversation was that rather than moving sensitive patient data to wherever the latest model lives, organizations should build a single, secure, well-governed data environment and then bring the models to that data. Bret noted that this approach can protect against both unnecessary resource churn and security risk, preserving up to 40% of the effort that would otherwise be spent re-securing data each time a model changes.

Is AI Actually Saving Money?

Bret was candid that the economics of AI at scale are still being worked out. The ROI calculus, he noted, is more nuanced than it first appears, and the industry needs new frameworks that account for the total cost of care, not just labor savings, to assess how to move forward with the technology in the most cost-effective manner.

He shared a real-world example in which the cost of LLM-driven automated coding began to approach what an organization had previously spent on offshore human coders, forcing a second look at where AI makes the most sense and where it doesn’t.

The ROI calculus is more nuanced than it first appears, and the industry needs new frameworks that account for the total cost of care, not just labor savings, to assess how to move forward with AI in the most cost-effective manner.

Intrinsic to this analysis, of course, is the fact that ROI is also not always financial. Aiding clinicians and staff in their work in such a way as to reduce burnout, ease patient engagement, and help healthcare organizations maximize their human resources is also a return on the investment in new AI technologies.

Training, Trust, and the Human in the Loop

I raised a question I think about often: what does user training look like when the underlying AI model is continuously evolving, rather than deployed and static for some period of time, as with most other health IT? Bret offered two angles.

On the technology side, governance must provide traceability, the ability to audit what changed in a model and why. For FDA-regulated devices, AWS has built capabilities to effectively “freeze” a model at a certified point in time.

On the user side, the more important change in focus is ensuring transparency and building trust by helping clinicians understand what data a model drew on and why it made a particular recommendation. That requires more frequent updates to users on how a model has changed, which is a training obligation that both AI developers and provider organizations need to embrace.

This reflects an encouraging regulatory trend on this front, with states like California and Colorado now codifying the “human in the loop” requirement into law, making clear that responsibility for AI governance is shared between vendor and provider.

This reflects an encouraging regulatory trend on this front, with states like California and Colorado now codifying the “human in the loop” requirement into law, making clear that responsibility for AI governance is shared between vendor and provider.

Testing AI in a Non-Deterministic World

An astute question from one of our members highlighted a fundamental tension: traditional software quality assurance is built on reproducible pass/fail behavior, whereas AI outputs can vary from run to run. How do you test against that?

Bret’s answer came back to traceability. When an AI output seems unexpected, the critical discipline is following the data trail back to its source to distinguish a hallucination or confabulation from a legitimate finding.

Regulatory frameworks for AI certification are still evolving, and approaches differ between federal and state governments, but that uncertainty doesn’t mean guardrails aren’t needed now.

Patient Safety and Privacy: Priority Zero

Bret was unequivocal on a final point: patient safety and privacy aren’t priority one; they’re priority zero.

…patient safety and privacy aren’t priority one; they’re priority zero.

He suggested that patient privacy is largely intact today because most AI workflows operate within established technology structures. In his opinion, the bigger ongoing concern is specific to patient safety: the risk that clinicians grow so comfortable with AI recommendations that they stop applying critical judgment, echoing familiar alert-fatigue dynamics. Many software developers, he noted, don’t yet have well-established processes for pre-market testing and ongoing efficacy validation. Both are non-negotiable given the stakes in healthcare.

Looking further out, Bret offered the personal observation that AI is poised to be the real accelerant for the consumerization of healthcare, giving patients far greater access to and control over their own health data than interoperability efforts alone have achieved. In that scenario, however, safety must still remain paramount in the design of technologies to be accessed by patients.

Closing Thoughts

In wrapping up our conversation, Bret and I agreed on this: healthcare is at a genuine inflection point in AI development and adoption. The organizations that come out ahead won’t be the ones chasing every new AI idea at once. They’ll be the ones that build secure data foundations, invest in meaningful clinical partnerships, prioritize a transparent governance model, and keep the patient squarely at the center.