and-healthcare.svg" alt="Tech and Healthcare" width="1200" height="800" loading="lazy" decoding="async">The intersection of technology and healthcare is one of the most consequential frontiers of our time. The promise is enormous: earlier diagnoses, fewer medical errors, more personalized treatment, lower costs. But every year I watch the gap between what's possible and what's actually delivered in a typical clinic stay roughly the same size.
Part of the problem is incentive alignment. Hospitals don't buy software the way a startup does. They have ten-year procurement cycles, regulatory overhead, and clinicians who, understandably, do not want to be guinea pigs for a buggy interface during a code blue. So innovation tends to settle in the edges — patient-facing apps, scheduling tools, billing — rather than the clinical core.
Where I'm more optimistic is in pattern recognition. A model that can read 50,000 dermatology images doesn't replace a dermatologist, but it can flag the four cases out of a hundred that genuinely need a second look. Same for ECG strips, retinal scans, mammograms. These are systems that augment rather than replace, and they tend to be the ones that actually ship.
The harder, more interesting work is downstream of the diagnosis — turning the right answer into the right action at the right time inside a real hospital. That is mostly a human problem dressed up as a technical one.
The Procurement Problem
One of the most under-discussed obstacles in healthcare technology is how hospitals actually buy software. Procurement cycles routinely run 18 to 36 months from first demo to signed contract. The buyer is rarely the user; the user is rarely the budget holder; the budget holder is rarely the IT integrator. By the time a product survives compliance review, security review, clinical-leadership review, and finance review, a competitor has often shipped a newer version. This is why so many promising healthcare startups stall after their first hospital pilot. The technology works. The sales motion does not.
The hospitals themselves aren't being irrational. A bad piece of software in a clinical workflow can hurt patients, and the regulatory consequences are severe. So caution is structural, not optional. But it creates a market where established vendors with mediocre products outcompete newer vendors with better ones, simply because the established vendor has already absorbed two years of someone else's pilot. The path forward isn't to lecture hospitals about velocity. It's to design software that survives the procurement gauntlet by being unambiguously safer than what it replaces.
Where AI Actually Helps
The most successful clinical-AI deployments I've watched have one trait in common: they augment a specialist who is already busy, on a task that is already standardized. Pattern recognition in mammography, ECG analysis, retinal screening for diabetic retinopathy — these are jobs where the model and the clinician disagree often enough that the second opinion is genuinely useful, and where the underlying images or signals are stable enough to train on at scale.
The deployments that fail tend to involve open-ended clinical reasoning. A model that promises to "predict patient deterioration" sounds wonderful in a press release, but the ground truth is messier than image classification, and the false-positive rate quickly desensitizes nurses. The real question isn't whether a model is accurate in a lab. It's whether the alert it generates changes the action a clinician would have taken anyway. If not, it's noise.
The Electronic Health Record Problem
Almost every conversation about technology in healthcare circles back, eventually, to the electronic health record. EHRs absorb an enormous fraction of clinician time and produce a correspondingly enormous amount of complaint. The interfaces are bad. The data models are baroque. The interoperability between systems remains poor a decade after federal mandates were supposed to fix it. Doctors routinely describe their EHR as the worst tool they use, and one of the leading drivers of professional burnout.
What makes this so frustrating is that the technical problems are almost all solved problems. Modern UI design, structured data exchange, sane access control — none of this is research. The problem is that the existing EHR vendors have decade-long contracts, deep integrations into hospital revenue cycles, and an effective duopoly on the largest health systems. Replacing them is closer to swapping out the foundation of a building than upgrading an app. Until that economic lock-in breaks, the rest of healthcare tech will remain partly stuck.
Patient-Facing Wins
The most visible recent improvements have happened on the patient side, not the clinical side. Online scheduling actually works at most large systems now. Test results reach patients via secure portals within hours. Telemedicine, accelerated brutally by the pandemic, is now a permanent option for huge swaths of primary and behavioral care. Insurance navigation tools, while still terrible, are measurably less terrible than five years ago.
These are real improvements, and they're being built largely outside the traditional hospital procurement loop. Direct-to-consumer healthcare companies, employer-sponsored benefits platforms, and pharmacy-anchored care models have moved faster than legacy systems because they don't have to negotiate with the EHR vendor every time they ship a feature. The next decade of healthcare tech will probably be defined less by hospital innovation and more by everyone else routing around it.
What I Keep Coming Back To
Whenever I get cynical about healthcare technology, I try to remember the specific case studies where it's already saved or extended lives — the diabetic retinopathy screenings that reach rural patients who would otherwise lose their sight, the sepsis-prediction models that buy clinicians an extra hour, the telemedicine consultations that get a child to a specialist their family couldn't have driven to. None of these systems are perfect. All of them are meaningfully better than what came before.
The system as a whole is moving slower than it should and faster than it ever has. Both things are true. The way to evaluate any specific technology is to ask whether it's pushing the average outcome forward, not whether it's solved the meta-problem of how healthcare is structured. The meta-problem will take decades. The individual wins are available now.