AI is everywhere in healthcare. Most of it is hype. Some of it is genuinely transformative. Here’s how to tell the difference — and where the real ROI lives.
Hamza Asumah, MD, MBA, MPH
I’ve sat through more AI pitches in the last three years than I can count.
Founders with impeccable decks and venture-backed confidence, promising that their platform will revolutionize diagnostics, eliminate administrative burden, and simultaneously improve outcomes and reduce costs by 40%. Sometimes all three at once.
Here is my honest assessment after all of that: maybe 20% of what I’ve seen actually moves the needle in real-world healthcare environments. The other 80% is sophisticated technology in search of a problem — dressed in the language of clinical improvement.
I want to be clear from the start: I am not anti-AI. I’m building AI-powered healthcare ventures myself, including platforms specifically designed for emerging market health systems. I believe deeply in the technology’s long-term potential.
But I am aggressively anti-hype. And in healthcare, where resources are constrained and the stakes of poor implementation are high, the gap between AI’s theoretical promise and its operational reality matters enormously.
How the Hype Cycle Works — and Why It Keeps Fooling Us
The AI hype cycle in healthcare follows a remarkably consistent pattern, and once you see it, you cannot unsee it.
A technology company publishes a study — often conducted under ideal conditions, on clean curated data, at an academic medical center with research infrastructure most healthcare operators will never have access to. The results are impressive. The study gets amplified in trade press and conference keynotes. Healthcare executives, under real pressure to demonstrate innovation to their boards, begin purchasing or piloting the tool.
The implementation reveals that the technology performs significantly worse in a real clinical environment than it did in the study. The data is messier. The workflow integration is harder. Staff adoption is lower than projected. The ROI doesn’t materialize.
The technology gets quietly deprioritized, the budget line disappears in the next fiscal year, and the vendor has already moved on to the next sales cycle with a new feature update and a testimonial from the one institution where it actually worked.
The operator’s test for any AI tool is simple: What specific outcome does it improve, by how much, and how was that measured — in an organization that actually looks like mine?
If a vendor cannot answer that question specifically and verifiably, you don’t have enough information to make a sound investment decision. Move on.
What AI Is Actually Delivering Today
Now let’s talk about what’s genuinely working — because the picture is not uniformly bleak. There are specific, high-value applications where AI is delivering real, measurable impact right now.
Diagnostic Imaging Analysis
This is the application with the most mature and credible evidence base. AI systems trained on large radiology datasets are performing at or near specialist-level accuracy for specific, narrow classification tasks — detecting certain cancers in screening mammograms, identifying diabetic retinopathy in fundus photography, flagging critical findings on chest X-rays.
The operative word is specific. These tools excel at well-defined, high-volume classification problems. They are not general-purpose diagnostic systems. The moment they’re positioned as such — or used as such — the clinical risk becomes significant. Used correctly, within their validated scope, they’re genuinely extending the reach of clinical expertise.
Revenue Cycle Management and Coding
This is where AI is delivering some of the most tangible and quantifiable ROI in healthcare operations — and it gets far less press than diagnostic applications because it’s not as cinematically compelling. But it matters enormously to the financial health of clinical organizations.
AI-assisted coding tools are reducing claim errors, accelerating prior authorization processing, identifying undercoding patterns that result in systematic revenue leakage, and flagging compliance risks before claims are submitted. In an industry where coding errors and claim denials cost healthcare organizations billions annually, a tool that measurably improves first-pass claim acceptance is not a minor efficiency gain — it’s a direct contribution to the bottom line.
Patient Communication and Retention Automation
AI-powered patient outreach — appointment reminders, recall campaigns, reactivation sequences for lapsed patients — is delivering measurable results in practices that implement it thoughtfully. This isn’t cutting-edge technology; it’s well-configured language models and automation pipelines. But the ROI is clear: higher appointment fill rates, reduced no-show rates, better patient retention, and more efficient front-desk operations.
In a five-location dental group, moving the no-show rate from 15% to 10% across all locations represents meaningful recovered production. If your average no-show represents $300 in lost production and you see 80 appointments per day across your group, that 5% improvement is $1,200 per day recovered. Do the annual math.
Operational Analytics and Pattern Recognition
This is the category I find most exciting from an operator’s perspective, and it’s the one I’m most actively building toward. AI tools that analyze your operational data — scheduling patterns, provider production trends, patient flow dynamics, AR aging trajectories — and surface actionable insights before problems become crises are genuinely changing how sophisticated healthcare operators manage their businesses.
Not by replacing human judgment. By ensuring that human judgment is applied to the right problems, with the right information, at the right time.
Where AI Consistently Fails in Healthcare
When the Underlying Data Is Poor
AI systems are only as good as the data they’re trained on and the data they’re applied to in production. Healthcare data is notoriously messy — inconsistent coding across providers and locations, fragmented records across systems that don’t communicate, years of data entry errors, and siloed information that’s never been unified.
An AI system deployed into a poorly organized data environment will not magically remediate that environment. It will confidently produce outputs based on bad inputs. Garbage in, garbage out is not a cautionary guideline in AI — it is a law of the architecture.
When Implementation Is Underestimated
Healthcare organizations — consistently, almost universally — underestimate the change management investment required to successfully implement AI tools. Clinicians and staff have to trust the output before they act on it. Trust requires demonstrated accuracy. Demonstrated accuracy requires trained and calibrated use. Trained use requires time, leadership commitment, and genuine workflow integration.
Most AI implementations fail not at the technology layer but at the adoption layer. The tool works adequately. Nobody uses it correctly. The ROI never materializes. The conclusion — incorrectly — is that AI doesn’t work in healthcare.
When It’s Positioned as Replacement Rather Than Augmentation
The fastest way to generate clinical resistance to an AI implementation is to frame it as a replacement for human judgment. Healthcare providers have strong, legitimate concerns about clinical accountability. The moment an AI tool is described — even informally — as ‘making the decision,’ adoption stalls.
The AI implementations that succeed in clinical environments are invariably the ones framed as decision support — tools that surface information, flag anomalies, and expand the clinician’s capacity. The clinician remains the decision-maker. The AI expands their information and efficiency. That framing is not just politically strategic; it’s clinically accurate.
A Framework for Evaluating Any AI Tool
Here is the evaluation process I use before recommending any AI investment in a healthcare organization.
Start With the Problem, Not the Solution
Write down your three most costly operational or clinical problems. Then ask, for each one: is AI actually the right solution? Often it is. Sometimes a better process, a rewritten workflow, or a different staff training program solves the problem more reliably and at a fraction of the cost. Don’t let vendor enthusiasm about their solution define what problem you’re trying to solve.
Demand Evidence From Comparable Settings
Comparable means similar size, similar specialty, similar payer mix, similar market. A case study from a 500-bed academic medical center operating in a large metro tells you almost nothing about what will happen in your eight-location multi-specialty group in a secondary market. Require proof from organizations that actually resemble yours.
Define Your Success Metric Before You Start
Before signing any contract, write down the specific, measurable metric that will define whether this implementation succeeded. Revenue per provider hour. First-pass claim acceptance rate. No-show rate. Treatment acceptance percentage. Commit to a number and a timeframe. If the vendor is unwilling to be evaluated against a concrete outcome metric, that tells you something important.
Pilot Before You Scale
Implement in one location with one team for 90 days. Measure rigorously. Learn. Adapt. Then expand. Healthcare organizations that roll out new technology system-wide simultaneously almost always generate chaos and distorted signal — making it impossible to determine whether the tool is the problem or the implementation is the problem.
The Operator’s Bottom Line
My framework for AI in healthcare is simple enough to fit in a sentence: if it doesn’t demonstrably increase revenue, reduce cost, or improve clinical outcomes — it is noise.
It might be impressive noise. It might be intellectually interesting noise. But if it can’t move one of those three needles in your specific organization, measured against a specific baseline, it is not a priority.
The best AI in healthcare is nearly invisible. It’s surfacing the right information at the right moment, flagging the right anomaly before it becomes a crisis, automating the right friction point. You barely notice it — but your outcomes are measurably better.
That’s the standard worth holding. And it’s exactly the standard I apply to every AI initiative I’m involved in building — including the platforms I’m developing for markets where the need for intelligent, resource-efficient healthcare technology is most acute.
The potential is real. The tools are improving every month. But the operator’s job is to stay grounded in business and clinical reality — not to follow the hype, and not to dismiss the genuine breakthroughs.
Discern carefully. Pilot rigorously. Scale what works.
Dr. Hamza Asumah 
Leave a comment