As Head of Engineering and Solutions at Binariks and a core member of their AI Center of Excellence, Mykhailo Hentosh has spent nearly two decades navigating the intricate architecture of healthcare software. His work focuses on the critical intersection of EHR interoperability, HL7 FHIR standards, and the deployment of agentic AI within real-world clinical environments. In an era where many healthcare organizations are struggling to move beyond experimental phases, Hentosh specializes in stripping away the hype to build systems that prioritize workflow integration and measurable ROI.
The following discussion explores the common pitfalls of hospital AI initiatives, such as the phenomenon of “pilot sprawl” and the structural failures that lead to high abandonment rates. We delve into the necessity of maintaining a “human-in-the-loop” philosophy, the technical friction of operating outside the patient chart, and a disciplined four-stage framework—Assessment, Pilot, Change Management, and Scale—designed to ensure AI solutions provide tangible value to clinicians rather than becoming expensive technical debt.
Many hospital AI pilots fail to deliver measurable ROI or result in “pilot sprawl” where active usage eventually fades into the background. From your perspective, why do these initiatives struggle to survive the transition from a demo to daily clinical practice?
The reality is that 95% of enterprise AI pilots fail to deliver any measurable ROI, and this failure is often baked into the project from the very beginning. We see “pilot sprawl” because organizations try to automate everything at once, creating a landscape of agents that no one can keep track of or justify. I’ve observed four main failure modes: the “Big Bang” where the budget is gone before a user sees value, “Pilot Purgatory” where a demo looks great but fails the reality of a messy clinical workflow, “Technology-First” builds where a solution hunts for a problem, and “Governance Paralysis” where red tape kills momentum. In a hospital, an annoying glitch that might be acceptable in a consumer app becomes a life-critical risk, so if the system feels opaque or untrustworthy, clinicians simply stop using it. They will always revert to the safest, most obvious manual decision if the AI doesn’t feel like a reliable partner in their actual environment.
You have mentioned that even a robust AI model can fail if it creates too much friction for the user. How does the choice between building an “orchestration layer” above the EHR versus integrating directly into the patient chart dictate the success of an AI agent?
This is a point of significant debate in the industry right now, but my experience over a decade of HL7 FHIR and EHR integration tells me that if an agent lives outside the chart, it is destined for the scrap heap. While some argue for separate orchestration layers to bypass the limitations of legacy EHRs, those layers often create a friction that clinicians won’t tolerate. A doctor shouldn’t have to leave the patient record, log into a second tool, and then toggle back just to act on an AI-generated recommendation. That extra step is where even the best triage engines go to die. To succeed, the agent must surface recommendations directly within the existing workflow using standard integrations, ensuring that the specialist can make a final clinical decision without breaking their focus or their digital workspace.
To avoid these pitfalls, you advocate for a disciplined sequence that begins with a four-week assessment. What specific metrics are you looking for during this initial phase to determine if a project is worth the investment?
The assessment phase is about ruthlessly reducing risk before a single dollar is spent on development. We spend about four weeks evaluating potential use cases against three core pillars: value, feasibility, and personal relevance. Value is the hard ROI, but feasibility determines if we actually have the data and infrastructure to pull it off today. Most importantly, personal relevance asks if the solution actually helps a specific human, like a floor nurse or a registrar, do their job better or faster. We produce a readiness scorecard at the end of this month, and if the project doesn’t hit the marks, we make a “no-go” call immediately. It is much better to stop a project in week four than to realize it was a mistake after six months of wasted effort.
Once a project clears the assessment, you move into a twelve-week pilot. How do you structure this “reality check” to ensure the AI remains practical rather than just a flashy technical exercise?
The pilot must be a real demonstrator working with live data and actual users in production, not just a slideshow or a sandbox demo. The golden rule I follow is that if you can’t explain exactly what the AI agent does to a floor nurse in one single sentence, the project is too broad and will likely fail. We focus on one narrow, clearly defined task where the system’s rationale is transparent; the specialist needs to see why a particular patient was flagged. During these twelve weeks, we monitor how the AI surfaces recommendations while keeping the human firmly in the loop for the final decision. When we built a documentation agent this way, we saw follow-up documentation time drop by about a third, and compliance with check-ins actually doubled, which mirrors the 27% reduction in documentation time seen in major health systems like Intermountain Health.
Change management is often the most neglected phase of AI deployment. How do you measure whether an AI tool is actually being adopted by staff, and what signals indicate that the system needs to be refined or pulled?
Change management is where technical excellence meets human behavior, and if you skip it, the project dies. We don’t just train the team and walk away; we look for simple, objective signals in the data to see if the tool has truly caught on. We track the ratio of how often a nurse accepts versus declines a suggestion, and what percentage of AI-generated recommendations are opened versus being ignored entirely. These aren’t just adoption metrics—they are a roadmap for improvement. If the ignore rate is high, it tells us the AI is likely “crying wolf” or surfacing information at the wrong time, and we use those insights to iterate on the workflow before we even consider moving toward a full-scale rollout.
As we look at the current landscape of healthcare technology, including federal transparency rules like the HTI-1 final rule, how should hospitals approach the transition from a successful pilot to a full-scale deployment?
Scaling should never happen just because “the pilot was well-liked”; it must be based on hitting predefined criteria like shift time saved, error rate reduction, and task compliance. With regulations like HTI-1 now demanding strict algorithm transparency in certified health IT, the era of “black box” AI is over. To scale effectively, you earn a slightly broader scope one use case at a time, moving away from the hype and toward evidence-based deployment. The hospitals that pull ahead in the coming years won’t be the ones trying to run twenty flashy experiments at once. Instead, the winners will be the organizations that run the fewest pilots but have the discipline to finish them and integrate them deeply into their operational DNA.
What is your forecast for the role of agentic AI in clinical workflows over the next two years?
I believe we are moving away from the “all-knowing” general AI models toward a modular ecosystem of highly specialized agents that do one narrow task exceptionally well. We will see these agents become invisible, meaning they won’t be perceived as “new software” but rather as a smarter layer within the EHR that handles the cognitive load of triaging, drafting, and flagging. The focus will shift entirely from the “intelligence” of the model to the “trustworthiness” of the integration. In this environment, the most successful systems won’t be the most ambitious or “intelligent” ones; they will be the ones that are easiest to govern, easiest to trust, and, most importantly, the ones that clinicians don’t even realize they are using because the friction has been completely stripped away.
