AI has moved into real healthcare operations. Ambient tools transcribe visits so clinicians can focus on patients instead of screens. Models read X-rays and retinal scans. AI increasingly generates the claims that get sent to insurers.
The hard part is rarely the model. It is getting the technology adopted, validated, and built into daily work. In a recent Playbook AI Partners podcast, Estenda Co-founder and COO, RJ Kedziora, laid out what makes AI stick across healthcare, MedTech, and Life Sciences.
How Do You Make AI Work Across Healthcare, MedTech, and Life Sciences?
Measure AI against how people actually perform
Ambient documentation is the use case healthcare adopted first and fastest. Clinicians, as RJ points out, "didn't get into healthcare to stare at an electronic medical record while they're talking to a patient." Letting AI capture the conversation hands that time back, which is why the accuracy debate around it matters so much.
Large language models draw steady criticism for hallucination and for being less than fully accurate. RJ does not dismiss that concern. He reframes the comparison.
"The pushback that I don't think is recognized enough is that us people, humans, we're not perfect either."
Medical scribes make the point concrete. Listening to visits and writing notes is a long-accepted role, and as RJ notes, "there are stats out there that the human scribes are not perfect. And if you implement a review process, those notes can be improved; they can be made better." AI note-taking earns trust the same way, by running through the review steps a health system already has in place. He describes the payoff from recent research as "a phenomenal amount of time" saved for physicians.
The same accuracy question is spreading past the exam room. AI now helps generate the billing sent to insurers, and RJ points to a near future where "an AI submitting the bill and the insurance company using AI" checks whether it is valid, which he calls "a whole nother world of challenges."
The standard stays constant across all of it.
Set the benchmark at real human performance, not perfection
Keep the review processes that already catch human errors
Track quality over time rather than expecting it on day one
That framing sits at the center of AI/ML development in healthcare, and generative AI is already tightening quality control in adjacent work, which RJ covered in how generative AI is improving healthcare software testing quality.
Keep clinicians in the loop as the validation layer
Some of the earliest FDA-approved AI use cases were in radiology, reading images and X-rays. We work in the same space, and RJ points to projects "for retinal, looking at retinopathy images, diabetic retinopathy, doing go-no-go decision making." Even there, a person never leaves the process. The AI is "speeding up the ability to read those images," but as RJ puts it, there is "still very much a human in the loop to validate what it's seeing."
That human check answers an early fear. Some predicted AI would "decimate the radiology industry" and that studying it was pointless. It did not play out that way, and RJ explains why. "We need those radiologists more than ever because they don't just read images. They provide so many other services and values."
For MedTech and Life Sciences teams, the pattern is a model that accelerates a specialist rather than one that stands in for the specialist. A retinal tool that speeds screening while a clinician confirms the read widens access without lowering the standard of care. The stakes climb as the clinician workforce ages and demand grows, and RJ frames AI as a way to "expand that capability" and reach more patients, which depends on preserving the human judgment those patients rely on. That balance runs through our diabetic retinopathy screening work and the validation studies that confirm a solution improves outcomes before it scales.
Lead with education and executive alignment
The first obstacle is hesitancy, and it is human. RJ sums up the questions people bring.
"Can we implement this AI solution? Is it the appropriate thing to do? Am I going to lose a job because of this?"
His answer starts with education, and it starts at the top. Executives, the board, and department leads have to understand what AI means and why it is being introduced so they can carry that message down. As he puts it, he would "start at the top" and "provide that education such that they can message the rest of the staff."
RJ names three keys in sequence.
Education: Teach leadership first so they can explain the change with assurance.
Alignment: Get the executive board and the technology staff agreed on the risks and the plan.
Change management: Message it across the organization and make it part of how people work.
He adds a reason to move deliberately without moving slowly. AI advances so fast that, as RJ describes it, a solution you cannot build today is one where you "just wait three months, it'll be able to do it." Experimenting to learn becomes part of the plan rather than a detour from it. A clear digital health strategy and roadmap turns that sequence into steps a whole organization can commit to, and the jump from approved idea to production system is broken down into what it takes to get a healthcare product to production.
Embed AI into existing workflows
A capable tool that sits outside daily work goes unused. RJ is direct about it.
"If it's not part of the workflow, it's never going to be adopted by the healthcare professionals or providers. It needs to be part of that workflow."
The obstacle now is data access, and the problem has flipped over twenty years. Where data was once scarce, it is now overwhelming, spread across medical record systems, journals, and wearables. RJ wears an Oura ring, a patient might wear an Apple Watch, and all of it has to reach the point of care so the AI can, in his words, "help you interpret that data and understand it." A person can now generate detailed data about their own health, bring it to a visit, and ask the doctor what it means, which only works if that data reaches the workflow in the first place. The work is connecting sources that were never built to talk to each other, including legacy EMR systems and streams of wearable data.
The systems can be connected, so what stalls projects is "operationalization," and as he says, "it's the people problem." Integration and interoperability are where the effort lives.
Estenda's custom software development work builds tools that connect to medical devices and fit the way teams already operate. The data analytics practice turns scattered data into something usable, and implementation and support get AI into daily practice without throwing off the workflow.
Build guardrails for accuracy, bias, drift, and privacy
RJ treats safeguards as the price of trust and names four that matter.
Data privacy: HIPAA obligations attach once a healthcare provider is involved, so institutions sign business associate agreements that extend those protections to the AI vendor. Data an individual enters into a consumer tool on their own, as RJ notes, "is not protected by HIPAA."
Bias: Models learn from human-made data and processes. As RJ puts it, "we as humans have biases," and that is "in turn reflected in the AI systems out there." Left unchecked, it can widen gaps in care, a risk covered in what it takes to build AI that includes underserved patients.
Drift: "These systems learn and change over time," RJ says, so monitoring keeps a model from "drifting away from where you think it's supposed to be operating."
Critical thinking: RJ calls this the biggest long-term risk. These tools are "designed to please you," and "they sound very confident," which makes it easy to accept an answer without checking. His guard is a habit of asking, "Is it really right? Is it really accurate?"
Regulation is still catching up to all four, a tension explored in is AI in healthcare moving faster than we can regulate it.
Ready to Operationalize AI in Your Organization?
Moving AI from idea to daily clinical use takes a partner who understands healthcare workflows, regulatory requirements, and the engineering behind safe, scalable systems. Estenda brings more than two decades of digital health, MedTech, and Life Sciences experience to that work, from strategy through deployment and support.
Book a free 30-minute consultation with a digital health architect to talk through your AI initiative, with no obligation. Reach out at info@estenda.com. Explore the full range of services at estenda.com.
Frequently Asked Questions
Is AI accurate enough to use in clinical settings?
AI does not have to be flawless to be useful. It has to perform at or above the people doing the task today, with review processes catching errors. Documentation and image reading already meet that bar in many settings when a clinician validates the output.
Will AI replace healthcare or technical staff?
Adoption has expanded roles more often than cut them. Radiologists were expected to disappear and became more essential, because they do far more than read images. The same holds in software, where AI writes code but people still direct it, secure it, and confirm it works.
How is patient data kept HIPAA compliant when using AI?
HIPAA applies once a provider is in the picture, so the key step is a business associate agreement that carries protections through to the AI vendor. Identifiable information entered into a consumer tool by an individual does not carry that protection, so institutions handle it carefully.
What is the biggest obstacle to adopting AI in healthcare?
It is usually a people and process challenge more than a technical one. Hesitancy, leadership alignment, change management, and connecting AI to legacy systems and workflows are where efforts stall. RJ describes this as an operationalization problem.
How long does it take to move an AI idea into a working product?
A working prototype can come together quickly. A production system that meets HIPAA and cybersecurity requirements and fits real clinical workflows takes longer and requires disciplined engineering. A defined roadmap keeps that path predictable.
What is model drift and why does it matter in healthcare?
Drift is how AI systems shift over time as they learn and as data changes. In healthcare, an unmonitored model can move away from its intended purpose and expected accuracy. Ongoing monitoring keeps performance aligned with clinical goals and patient safety.
