Healthcare AI Is Solving the Wrong Problem, and We’re Calling It Progress

Billions of dollars have flowed into healthcare AI over the last decade, and the results are, by many conventional measures, impressive. Documentation is cleaner. Billing cycles are shorter. Administrative throughput has never been more efficient. But a more important question still lingers. Why does the patient experience still feel so utterly fragmented and reactive?

It’s because we built the machine for the institution, not the person.

Healthcare AI is serving reimbursement cycles, compliance frameworks, and operational throughput. But these are metrics that serve the system’s financial architecture, not the human being inside it. They don’t compensate for the lack of care. Incentives shape outcomes, and right now, those incentives are misaligned with what it actually means for a person to get better.

Speed has become a proxy for progress, but it’s hardly doing any justice to the meaning of advancement. While it is easy to measure, when applied to a flawed system, it only accelerates the flaws. Automating inefficiency at scale does not transform care, but reinforces the very structures that limit it.

I keep coming back to the data. More than 80% of physicians now use AI in some form, a figure that has more than doubled in recent years. That shows progress, doesn’t it? Except that the depth of use remains narrow, reducing its potential for administrative work. The tasks that AI has made faster are, by and large, the tasks that matter least to whether a patient actually gets better. They are certainly useful, but they do not fundamentally change care delivery. They optimize the periphery while leaving the core untouched, leaving us to celebrate the wrong wins.

And that misalignment goes deeper into what we choose to measure. Healthcare AI is largely built on structured data because structured data is convenient. Labs, vitals, diagnostic codes, these are clean and standardized. They fit neatly into models. But the earliest signals of a patient’s decline rarely live in those numbers. Those indications show up in behavior, in routine changes, in oscillating moods, and in gradual withdrawal.

An erosion of the small daily patterns, when tracked over time, tells the real story of how someone is doing.

These signals are messy and continuous, often ambiguous. As a result, they can get ignored entirely, leaving models to only learn what is easy to capture instead of what is important to understand.

What does this mean in practice? It means we have built models that are extraordinarily precise about what they measure and essentially indifferent to everything else.

I often say healthcare doesn’t have an intelligence problem; it has a memory problem. Every interaction resets the narrative. A patient sees a provider, shares part of their story, and leaves. The next encounter starts again with partial context. Information then becomes lost between visits and between systems. AI has inherited this limitation. It’s operating on snapshots, not the full story, and in the end, a prediction made without longitudinal context is, at best, statistical noise.

The industry may respond by generating more alerts, more flags, more signals, each competing for attention. Yet that inevitably leads to alert fatigue, which again, isn’t a usability issue, but it is a design failure. Systems do not understand what matters, so they surface everything, and when everything demands attention, nothing is truly prioritized.

Meanwhile, the factors that drive outcomes remain invisible within these systems. They sit outside the formal data structures. We call them “soft signals,” yet they often carry the hardest consequences.

Intervention arrives late as care is triggered by events, such as a hospitalization or a measurable deterioration. These outcomes trace back, more often than not, to the invisible factors, like loneliness and behavioral drift. Research consistently identifies social determinants among the strongest predictors of mortality and diseases, yet they sit almost entirely outside the scope of what most healthcare AI is designed to capture. 

I see it as a missed opportunity. True progress would involve understanding earlier, before escalation occurs. Prevention requires something the current architecture cannot provide: continuity. A thread of understanding that connects who the patient was yesterday, to who they are today, to who they are becoming. Without that thread, you are just guessing more quickly.

There is also a troubling dynamic in how healthcare AI gets adopted in the first place. Boards and executives feel the pressure to have an AI strategy, so the question shifts from “What problem are we actually solving?” to “Where can we add AI?” I believe that inversion explains why so many pilots stall.

The demos are impressive, but the real-world impact? Quiet and absent.

Fragmentation compounds this problem. Healthcare generates vast amounts of data, yet coherence is elusive. Information lives across systems that do not communicate effectively, and AI models get trained on what is available within those silos.

In that process, an abundance of data is paired with a scarcity of understanding. Critical context gets missed, and decisions get made as though it were present.

Adoption challenges reinforce this. Tools, essentially designed for executives and administrators, are often used by clinicians under significant pressure. This can cause priorities to diverge. So if a solution doesn’t integrate seamlessly into existing workflows, it effectively doesn’t exist in practice. Usability, trust, and relevance can determine whether technology becomes part of care or remains an outlying experiment.

I don’t believe the future of healthcare will be defined by better predictions alone. Prediction has value, but without understanding, it remains incomplete. The real shift will come from systems that connect data, context, and time as a living, longitudinal model of the individual. One that makes the invisible visible by encompassing the behavioral shifts, the social patterns, and the ambient signals that precede every crisis.

Until healthcare has memory, it cannot have intelligence. And until it has intelligence oriented around the person rather than the institution, it will keep solving the wrong problem at an ever-increasing speed.

About the Author

Freddy del Barrio is a founder building critical infrastructure for essential, underserved industries. He is the founder of Companion AI, a platform dedicated to scaling healthcare infrastructure through intelligent systems that enhance operational efficiency, staff support, and patient engagement.

Freddy works across healthcare technology, AI deployment, and private markets. He focuses on empowering large care networks and healthcare platforms through the integration of strategic capital and transformative technology.