Health AI has an accuracy obsession. Models are benchmarked, papers are published, and performance metrics climb each year. Yet inside the clinic, adoption tells a humbler story. For Nihit Gurram, founder of Mosaic Health Solutions and a health informatics graduate student building clinical decision support for medication safety in older adults, the industry has been optimizing the wrong variable.
“Accuracy is table stakes,” Gurram says. “The real question is whether a clinician with fifteen minutes, a full waiting room, and an already crowded screen will use the tool at all. If the answer is no, the model’s performance is irrelevant. Adoption, not accuracy, is the bottleneck, and it has been for years.”
Designing for the Fifteen-Minute Visit
Gurram’s design philosophy starts with a constraint most health technology treats as an afterthought: the clinical visit is short, and every second of it is contested. He calls the resulting principle the fewest-clicks rule. Decision support has to fit inside the visit rather than fight it, surfacing insight in the time it takes to glance at a screen, not the time it takes to open one.
That leads directly to workflow-first design. “The tools that get used are the ones embedded where the clinician already works,” he explains. “The moment you ask someone to open another tab, log into another system, or break their rhythm mid-visit, you have lost them. Not because they are resistant to technology, but because the workflow is the job, and anything that interrupts it costs patient time.”
Latency as a Clinical Feature
One of Gurram‘s more contrarian positions is that response speed deserves to be treated as a clinical feature, not an engineering detail. A slow tool, in his framing, is functionally a broken one, because clinicians will route around anything that makes them wait. The same logic applies to data fragmentation. When patient information sits in silos, no single view captures the full phenotype, and the clinician is left assembling a picture the system should have assembled for them. Un-siloing that data, he argues, is as much an adoption problem as a technical one.
The Cost of Crying Wolf
Then there is alert fatigue, the quiet killer of clinical software. Systems that flag everything teach their users to dismiss everything, and each false alarm spends down a finite budget of attention. Nihit Gurram believes the discipline of designing for signal over noise matters most in exactly the domain he works in, medication safety for older adults, where the volume of potential interactions could justify an alert on nearly every chart.
“If your system cries wolf, clinicians will silence it, and they will be right to,” he says. “The hard work is deciding what not to surface. Restraint is a feature.”
Closing the Trust Gap
Underneath all of it sits what Gurram calls the trust gap. Clinicians are trained to justify their reasoning, and they extend the same expectation to the tools advising them. A recommendation without a rationale is a liability, not an assist. This is why he builds what he describes as glass-box AI, systems that show their reasoning and cite the established clinical criteria behind each flag, keeping the clinician in control rather than asking for blind trust.
“A black box asks the clinician to outsource judgment,” he says. “A glass box earns its way into the workflow by making the clinician’s own judgment faster and better supported. In medicine, showing your work is not a nice-to-have. It is the price of admission.”
Gurram is careful to frame his perspective as that of a technologist and incoming medical student at the Kansas College of Osteopathic Medicine, not a practicing physician, and his conclusions are about design rather than clinical practice. But the throughline of his work is hard to argue with: in health AI, the systems that win will not be the ones with the best benchmarks. They will be the ones clinicians actually reach for.