gshc2020.com

But somewhere along the way, the data stopped being the bottleneck. It is every human being still sitting between the data and the decision it's supposed to inform. A population health nurse gets a risk-stratified list every morning and still spends hours triaging it against her own read of the panel because nobody has rebuilt the workflow to trust it. A prior auth coordinator gets a decision-ready determination, complete with clinical rationale, and picks up the phone to call the payer anyway. That's the workflow. It has always been the workflow, and nobody rebuilt it just because the recommendation got smarter.

None of this is a people problem, and I want to be clear about that because I've watched this argument get twisted into “healthcare workers aren't trying hard enough,” which is exactly the opposite.

Healthcare has invested heavily in trying to fix this, and adoption is no longer the obstacle people assume it is. Two out of three physicians now use AI tools in their practice, according to the American Medical Association, up sharply from just a couple of years ago. And yet administrative costs, according to the American Hospital Association, now consume an estimated 25% to 35% of every dollar spent on healthcare in this country. Medical prior authorization, one of the most studied examples, is still processed electronically only about 40% of the time, despite two decades of automation investment aimed squarely at fixing it. The honest question is why, when AI is deployed at this scale, it hasn't changed the underlying number.

Most of this AI is being added on top of an architecture that was never designed to enable a model to act. Healthcare doesn't have an intelligence problem anymore. It has an architecture problem, and until that gets rebuilt, better AI will keep producing the same disappointing math.

A Closer Look At The Problem

Most AI in healthcare today flags. It suggests. It surfaces. A human still reviews it, still approves it, still does the work, just a little faster than before. But it changes nothing about how the workflow is built, so it inherits every bit of friction the old workflow had. You've made someone quicker at a task.

Digitization did the same thing a generation earlier. We replaced paper charts with EHRs and filing cabinets with dashboards, and somehow kept the same number of people translating between systems that still can't talk to each other directly. Thirty years of this. If better tools layered on top of the old structure were going to fix it, we would have seen it by now.

What A Rebuilt Foundation Looks Like

What's actually missing isn't a smarter model. It's the plumbing that lets a model act on what it concludes, in real time, on complete data, and learn from what happened next. Most of healthcare AI stops at recommend, and you can see exactly where an organization sits by asking how far past that word it's actually gotten:

Level 1: Visibility
People see everything and still decide and act on it all.

Level 2: Assistance
AI recommends. Humans still decide. Most health systems are here.

Level 3: Partial Automation
Tasks automate in silos. Humans manage the handoffs. The plateau.

Level 4: High Automation
Workflows run end-to-end. Humans handle the exceptions. The realistic goal.

Level 5: Fully Autonomous
Continuous, end-to-end execution. Humans oversee governance, not workflow.

Most health systems are stuck at L2 or L3. Closing that gap is architectural, not a feature you buy.

I'll say the thing that usually gets left out of these arguments: Less than 15% of the administrative work in revenue cycle, care management and utilization review actually requires human judgment. Fix the architecture and you don't remove the nurse or the coder from healthcare. You remove them from the eight-day approval loop and give them back the hours to do what they went into this field to do. Caring for the patients.

So what does rebuilding actually take? Not a bigger AI budget, and not another vendor evaluation. It starts with leaders asking a different question for every workflow they run: Does this step exist because human judgment is genuinely required, or because nobody has built the plumbing to run it without one? The pattern is consistent enough that I'd call it a law: Wherever a system was designed around a human checkpoint by default, adding a smarter tool never removes it. Only redesigning the workflow does.

The organizations willing to have that conversation are already seeing the results. Claims that used to take a week get resolved the same day. Prior authorizations that required a phone call clear without one. Nurses spend their mornings on patients who genuinely need a human judgment call, not reverifying a list the system already got right. That's what happens when you stop asking software to be faster inside the old structure and start asking it to run the parts that never needed a person. The organizations that treat this as an architectural problem are going to end up somewhere very different from those still buying point solutions.

The only path forward I can see is healthcare autonomy: building systems where the work that shouldn't require a person doesn't get one, so the people who came into this field to care for patients can finally spend their time doing that. The tools to build it exist now. What's missing is a stronger will to rebuild the foundation rather than decorate the one we have.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?