4% of Health Systems Have Cracked AI ROI. Here's What They Did Differently.

There's a number that stopped me when I read it last week.

Becker's reported that only 4% of health systems have achieved scaled AI implementation with measurable outcomes. Four out of one hundred. And four out of five health system leaders said they have difficulty measuring AI ROI. This is after years of investment, conference keynotes, pilot programs, vendor contracts.

Ninety-six percent are still stuck.

I find this clarifying. Because I've watched the 4% do their thing, and I've watched the 96% do theirs. And the difference isn't the AI. It's everything around the AI.

What the 4% Actually Did

CommonSpirit Health: $100 million in AI value in a single fiscal year. 242 applications live. They also rejected 15 proposed use cases through their governance process. That last part is the tell.

Penn Medicine: projecting $105 million in AI benefits by 2028. 20% productivity improvement among project managers. Clinical tools in radiation oncology. They call it "innovation with guardrails."

Mount Sinai: $50 million bottom-line impact projected this year. 3:1 ROI. Their strongest example is pressure injury prevention, not the flashiest use case, but one with clear measurement.

What do these have in common? None of them deployed AI without knowing how they'd measure it. They built governance frameworks first. They defined what success looked like before the first dollar was spent. They had people who could say no to bad ideas.

That's boring. That's also how you go from 4% to the majority.

Why the 96% Are Stuck

The 96% aren't stupid. They hired the right consultants. They went to the right conferences. They read the same white papers. They signed the same vendor contracts. But they deployed AI the way most organizations deploy any new technology: they bought the tool and hoped the tool would solve the problem.

The tool doesn't solve the problem. The workflow solves the problem. The measurement strategy solves the problem. The willingness to kill projects that aren't working solves the problem.

Four out of five health system leaders saying they "have difficulty measuring AI ROI" isn't a measurement problem. It's a design problem. They didn't build measurement into the project from the start. So when the vendor's dashboard shows them 50,000 AI interactions and they have to explain to the CFO what that means. They can't.

I've been in that room. Not as the health system leader. As the AI vendor. And I can tell you: the conversations that go well are the ones where the health system came in with a baseline. A readmission rate. A no-show rate. An engagement rate. A cost per avoided admission. The conversations that go nowhere are the ones where we're talking about "transformation" without a number attached to it.

What This Means for Healthcare AI Infrastructure

The Becker's piece is really about infrastructure. Not the AI itself. The underlying operating layer that makes AI deployable, measurable, and scalable across a system.

CommonSpirit's 242 live applications don't exist because they found 242 great AI vendors. They exist because they built the pipes. Governance. Integration standards. Measurement frameworks. A team that could say yes or no with conviction.

This is what separates patient engagement AI that works from patient engagement AI that sits in a dashboard nobody looks at. The technology at the edges is roughly equivalent across vendors in 2026. What isn't equivalent is whether the technology connects to the EHR, whether the escalation pathway is defined, whether the clinical team knows what to do when the AI flags something, whether someone owns the readmission metric.

At HANA, we're self-hosted and open-source, which means you own the infrastructure. The data stays with you. The integration is yours to control. We've deployed in five countries, three languages, over a million interactions. The 85% weekly engagement we see isn't magic. It's the result of structured outreach, defined escalation, and clinical teams who know what to do with the signal. And the 31:1 ROI we document comes from knowing what the baseline was. You can read the case studies and the research that back those numbers.

How Do You Get From 96% to 4%?

You start with one use case. One metric. One baseline.

Not "AI strategy." Not "AI transformation." One clinical workflow where the current outcome is measurable and suboptimal. Post-discharge follow-up: what's your 30-day readmission rate? Medication adherence outreach: what's your six-month A1C change in the relevant population? Care gap closure: what percentage of patients with elevated blood pressure had a follow-up within 30 days?

You run AI on that workflow. You measure the same metric. You compare.

The health systems in the 4% didn't do anything exotic. Mount Sinai's best example is pressure injury prevention. CommonSpirit's headline number comes from 242 individual projects, each with its own measurement. Penn Medicine's team tracks productivity improvement among project managers.

These are not moon shots. These are operational improvements with a number attached.

The ones still stuck are waiting for the transformation to be obvious. It never is. ROI in healthcare AI accrues through accumulation: small improvements, measured consistently, compounded over time. The health systems that got there built the infrastructure to see it.

What We're Seeing in the Field

In HANA deployments, we see a pattern that matches what Becker's is describing. The deployments that produce measurable ROI quickly are the ones where the health system already has a baseline readmission rate, a defined escalation pathway, and clinical ownership of the metric. The deployments that stall are the ones where AI gets handed to IT as a "technology project" rather than owned by the clinical team with the most to gain.

The 4% treated AI as a clinical operations problem. The 96% treated it as a technology problem.

That's it. That's the whole difference.

My daughter doesn't care about our AI strategy. She cares whether patients get better. The 4% care about that too. They built systems to prove it.

The Infrastructure Question Is Now Unavoidable

Healthcare AI in 2026 is not in the evaluation phase. The Becker's data makes that clear: 70% of healthcare organizations are actively using AI. See how HANA is priced for this model. The question is no longer whether to deploy. It's whether you're in the 4% who can measure it, or the 96% who can't.

The infrastructure that makes measurement possible isn't glamorous. It's data standards, integration work, governance committees, defined escalation pathways, owned metrics. It's the stuff that doesn't make conference keynotes.

But it's what separates the health systems posting real numbers from the ones still talking about transformation.

Build the pipes. Define the metric. Run the project. Measure the result.

That's the whole playbook. And it's available to anyone who decides to use it.

If you want to see how HANA fits into this, book a discovery call. We'll walk through your current metrics, your integration environment, and what a pilot would look like with a defined outcome from day one.

Key Takeaways

Only 4% of health systems have achieved scaled AI with measurable outcomes, per Becker's Hospital Review. Not because the AI is bad. Because most organizations skipped the infrastructure that makes measurement possible. The health systems in the 4% built governance frameworks, defined metrics before deployment, and maintained ownership of clinical outcomes. HANA's deployment model is self-hosted, open-source, and EHR-integrated: built for exactly this. The 85% engagement rate and 31:1 ROI we document exist because every deployment starts with a baseline and a defined escalation pathway. The same model available to health systems is available to any practice that wants to close care gaps at scale.

FAQ

Why are so many health systems struggling to measure AI ROI?

According to Becker's, 80% of health system leaders say they have difficulty measuring AI ROI. The core issue is that measurement wasn't built into the project design from the start. Organizations bought AI tools without establishing baselines, defining success metrics, or creating ownership for the clinical outcome they were trying to improve. Without a pre-deployment readmission rate or engagement rate to compare against, any post-deployment number is meaningless.

What's a realistic starting point for a health system that wants to get into the 4%?

Pick one workflow with one measurable outcome. Post-discharge follow-up is a strong first use case: 30-day readmission rates are well-documented, the intervention is well-understood, and the clinical value is established. Start with a 90-day pilot on a defined patient population, document the baseline, run the program, and measure the same metric at the end. That's how CommonSpirit, Penn Medicine, and Mount Sinai all started.

Does being open-source and self-hosted actually matter for ROI?

It matters for sustainability. If your AI vendor disappears or changes pricing, a proprietary black-box system creates dependency that's expensive to unwind. Self-hosted infrastructure means the data, the integration work, and the model stay with you. For health systems building toward scaled deployment (the kind that produced $100M in value at CommonSpirit), data portability and infrastructure ownership compound over time. For smaller practices, it means lower total cost and full control over PHI.