Use AI as an Orthotic, Not A Prosthetic

Here's a question nobody in healthcare wants to answer honestly: If your AI strategy is built around replacing people, what happens when the algorithm gets it wrong?

Because it will get it wrong. The data tells us that. And in healthcare, "getting it wrong" doesn't mean a bad product recommendation. It means a missed diagnosis. A wrongful denial. A patient who falls through the cracks.

After 22 years leading quality improvement inside health plans and provider groups, navigating CMS audits, Stars ratings, and HEDIS cycles, I've watched this industry adopt technology the wrong way more times than I can count. The pattern is always the same: buy the tool, cut the headcount, celebrate the savings, then scramble when the cracks show.

There's a better framework. It comes down to one question: Are you building a prosthetic, or are you building an orthotic?

Prosthetics Replace. Orthotics Support.

A prosthetic replaces something that's missing. It takes the place of a limb. In healthcare AI, a "prosthetic" strategy means you're trying to automate away the human: replace the coder, replace the reviewer, replace the nurse's judgment with an algorithm. Cut the headcount, save the money.

An orthotic is different. An orthotic is a brace. It supports something that's already there. It aligns, it corrects, it stabilizes. It lets the body do what it's designed to do… just better, with less pain, and with less risk of injury.

When I look at the AI landscape in Value-Based Care right now, I see an industry overwhelmingly building prosthetics. And the evidence says that's a mistake.

The Data: AI Alone Gets It Wrong Nearly Half the Time

A white paper from Chirok Health studied AI in HCC coding (Hierarchical Condition Category coding), the financial lifeblood of Medicare Advantage risk adjustment. Get it right, and your plan captures accurate revenue. Get it wrong, and you're either leaving money on the table or inviting a CMS audit.

They studied roughly 3,700 patient records. The results were striking.

Under the prosthetic model, where AI worked alone, it identified 0.73 conditions per patient. Clinical reviewers only agreed with 54 percent of the suggestions. Almost half were wrong or needed modification.

Under the orthotic model, where AI was paired with human reviewers, the team identified 1.06 conditions per patient. That's a 45 percent improvement in productivity. And when those human corrections were fed back into the AI, accuracy jumped another 17 percent.

This isn't a story about AI being bad. It's a story about AI being incomplete. A scalpel without a surgeon is precise, sharp… and dangerous.

Key Finding: AI alone: 0.73 conditions/patient at 54% accuracy. AI + human reviewers: 1.06 conditions/patient, a 45% productivity gain, with a 17% accuracy improvement from feedback loops.

Three Ways AI Fails Without Human Oversight

The Chirok team categorized the errors into three types. Once you see them, you'll recognize them everywhere.

1. Errors of Omission

The AI simply couldn't read certain parts of the medical record: hand-written notes, information stored in unusual EHR locations. A human reviewer catches this in the first thirty seconds of opening a chart.

2. Errors of Interpretation

The AI could see the data but couldn't apply it correctly. For drug and alcohol dependence (HCC 55), 46 percent of AI suggestions were overturned by information in the medications list. There are four commonly prescribed medications for substance use disorder. A human reviewer sees those medications and immediately makes the connection. The AI couldn't do it consistently.

3. Errors of Conceptualization

This is the most dangerous failure mode. The AI couldn't sequence clinical events into a narrative. It couldn't tell the difference between a "rule-out" note and an active diagnosis. Twenty-three percent of all errors came from the AI misinterpreting progress notes. The AI could scan data points, but it couldn't read the story.

"The Human Body Is a Black Box"

This connects to a broader truth about healthcare. A research team at Duke Health built one of the first deep learning models fully integrated into routine clinical care: Sepsis Watch, designed to predict sepsis, the leading cause of inpatient deaths in U.S. hospitals.

One of their team members made an observation that should be tattooed on every AI vendor's pitch deck: the human body is a black box. Not the algorithm. The body. Sepsis doesn't even have a universally accepted definition. When human experts from across the country reviewed the same patient cases, they frequently couldn't agree on the diagnosis.

So when someone tells you their AI has "solved" clinical decision-making, ask them how they solved a problem that the world's leading clinicians can't even define consistently.

The Sepsis Watch Blueprint: How Duke Did It Right

The Duke team didn't throw up their hands. They built the tool, but they did it the orthotic way.

They started with the problem, not the technology. Front-line physicians submitted a proposal to improve sepsis detection. The project wasn't driven by a vendor pitch or an innovation lab looking for a use case.

Then they designed a workflow that respected professional discretion. When the AI flagged a high-risk patient, it didn't fire an alert directly to the ER doctor. A previous attempt had done exactly that: a pop-up firing over 100 times per day, with 86 percent of notifications canceled. Classic alarm fatigue.

Instead, they created what I call a "watchtower model." The AI monitored patients in the background. When it flagged someone, a Rapid Response Team nurse reviewed the patient's chart, synthesized digital data with clinical context, and called the physician. Not with an alarm; with a conversation: "I'm seeing a risk spike here. What do you think?"

The AI was the brace. The nurse was the body. The physician made the final call.

And here's the part nobody anticipated: the nurses developed entirely new expertise. They became skilled at remotely evaluating sepsis risk by synthesizing digital patient data, a capability that didn't exist before the tool was deployed. The technology didn't replace their expertise. It created new expertise.

The Orthotic Principle: Technology didn't replace expert judgment at Duke. It elevated it. Nurses developed new capabilities. Physicians retained final authority. Patients got better care. That's the orthotic model.

The Regulatory Reckoning Is Coming

CMS is paying attention. Researchers at Stanford have described what's happening in health insurance utilization review as an "AI arms race." Payers automate prior authorizations. Providers automate appeals. Both sides accelerate. And when you automate "no" at scale, you risk wrongful denials at scale.

CMS has signaled clearly in the Contract Year 2026 Policy and Technical Changes that they are closing loopholes in Medicare Advantage appeals. The scrutiny is coming, and AI-driven decisions that can't be explained, traced, or audited will be the first casualties.

In a Value-Based Care world where we are judged on patient outcomes and Star Ratings, we cannot afford black-box denials. We need AI that helps human reviewers make better decisions, not just faster ones.

Five Principles for Building the Orthotic Enterprise

A major review from Vanderbilt, Stanford, Emory, and UAB proposed a lifecycle framework for healthcare AI that every leader in this space should study. Combined with the lessons from Duke and Chirok, here's the framework I'd recommend:

Start with the problem, not the technology. Do not begin with "We need AI." Begin with "What is the specific clinical or operational problem we're trying to solve?" The Duke team succeeded because front-line physicians defined the problem. The Chirok study showed that when clinicians guide AI, outcomes improve dramatically.

Clinician-in-the-loop is non-negotiable. Whether it's generative AI for drafting patient messages, predictive AI for risk stratification, or analytical AI for Stars performance, a human must verify the output. The orthotic only works if someone is wearing it.

Explainability builds trust. If your team doesn't understand why the AI is flagging a patient or prioritizing a measure, they'll ignore it. The Duke team created "Model Facts" labels; essentially nutritional labels for algorithms. Transparency isn't a nice-to-have. It's the foundation of adoption.

Build feedback loops, not static tools. When the Chirok team fed clinical reviewer corrections back into the AI, accuracy improved by 17 percent. When the Duke nurses identified issues, the model evolved. Your front-line staff are your best calibration engine.

Manage the human side of change. AI shifts how people work and can feel like surveillance if deployed without empathy. The Vanderbilt framework emphasizes that human well-being and fairness must be central to implementation, not afterthoughts. The best technology in the world fails if your people don't trust it.

The Bottom Line

If your AI roadmap is built around replacing headcount, you're building prosthetics. You're inviting regulatory scrutiny. You're scaling errors of omission, interpretation, and conceptualization. And you're failing to tap into the most powerful pattern-recognition engine in healthcare: the trained human professional.

But if your roadmap is built around augmenting decision-making, reducing administrative burden, and closing care gaps… you're building orthotics. You're supporting your people so they can focus on the patient. You're building the kind of transparent, traceable, human-centric AI that regulators reward and clinicians trust.

In Value-Based Care, our currency is trust. Trust between the payer and the provider. Trust between the doctor and the patient. Trust between the algorithm and the human reading its output.

AI used as a prosthetic erodes trust. It feels like a machine saying "no."

AI used as an orthotic builds trust. It feels like a safety net ensuring the right "yes."

Let's build braces, not replacements.

This post is adapted from Episode 1 of The Orthotic Mindset, a podcast about stabilizing Value-Based Care with human-centric AI. Hosted by Berto Rico, 22-year healthcare quality veteran and Founder & CEO of ClearStars.AI.

Follow the show and share with a colleague navigating AI strategy.

Sources:

• Chirok Health LLC. "Evaluating the Role of AI in Value Based Care Premium Adjustment." White Paper, October 2024.
• Sendak, M. et al. "The Human Body is a Black Box: Supporting Clinical Decision-Making with Deep Learning." FAT* '20, 2020.
• Al-Garadi, M. et al. "Large Language Models in Healthcare." Vanderbilt, Stanford, Emory, UAB. 2025.
• CMS Contract Year 2026 Policy and Technical Changes to the Medicare Advantage Program.