Three Years of AI in Healthcare: Efficiency vs Transformation

Three years ago—hard to believe—I wrote my first article on AI in healthcare, right as OpenAI launched ChatGPT. I discussed where I thought generative AI would have the biggest impact on healthcare and medicine. I vividly remember writing it. The possibilities felt euphoric. I could envision how care might be transformed.

Now, three years later, it's a holiday weekend and no one wants to read about health policy or drug pricing. So I thought I'd revisit that first article. The question: are we actually implementing AI in the areas I thought would matter most?

Healthcare Implications of AI Revisited

I initially organized AI's healthcare implications into three main buckets:

Patient care and delivery
Research, diagnostics, and treatment
Clinical and non-clinical workflow

Patient Care and Delivery

I envisioned AI improving patient care and delivery through triaging, detecting errors, and engaging patients in their care plans.

Three years later, this is where I've been most surprised, not because AI hasn't made progress, but because the progress exposed how much of the problem was never technical in the first place.

Triaging and access turned into a $100+ million market that grew 20x year over year. AI platforms like Counsel Health, Doctronic, and Torch Health now assess symptoms conversationally and route patients to appropriate care. Scheduling automation tools like Assort Health and Hello Patient eliminate manual appointment booking. This was the space I tried to crack with Nayroo—matching patients to open slots and reducing no-shows.
Patient engagement is where the fun and real transformation happened. AI care navigation platforms like Hippocratic AI, Ferry Health, and Aidify now handle what we used to abandon patients on: the dead zone between clinical encounters. They call with results. They schedule follow-ups. They coordinate care transitions. They answer questions at 2 AM. Even within my own discharge workflow, I’ve incorporated using DoximityGPT. I now routinely convert my physician-directed discharge summaries into patient-friendly hospital course explanations and warm handoff letters for PCPs—both in under a minute. Patients actually read them and family members reference them. PCPs get actionable information during the highest-risk post-discharge window instead of operating blind.

One caveat with all of these tools, though. Most of them are efficiency AI, not transformative AI. They make broken processes faster. They help us navigate dysfunction more smoothly. They don't fix the dysfunction itself. We're automating abandonment instead of eliminating it.

Research, Diagnostics, and Treatment

The diagnostic and clinical decision support space evolved differently than I thought. Instead of AI replacing clinical reasoning, we got AI-augmented medical references.

UpToDate launched Expert AI that emulates how expert clinicians reason through problems, with full transparency into assumptions, sources, and reasoning steps. OpenEvidence raised $210M for rapid literature synthesis. Doximity acquired Pathway Medical for $63M, integrating it into DoxGPT. ClinicalKey, DynaMed, AMBOSS, and Glass Health all launched AI-powered clinical decision support.

What matters is that the answers are transparent, grounded in trusted content, and designed to augment rather than replace clinical judgment. This transparency matters enormously for adoption.

If physicians and trainees use these tools correctly—that is, with intent to learn and improve their knowledge base—then I do think the next generation of physicians will be some of the brightest ever. I’m biased, though.

Clinical and Non-clinical Workflows

I’m big on using AI to reduce administrative burden, streamline documentation, and eliminate unnecessary work. This is where the AI arms race is happening, and also where I've become most skeptical.

AI scribes are now ubiquitous. They're the "pager" of healthcare now: everyone has one, and if you don't, something's wrong. Abridge, Suki, Nabla, Doximity and dozens of others automate clinical documentation. That's valuable. It buys physicians time. But it doesn't question why we're generating so much documentation in the first place.
Prior authorization AI exploded as spending grew 10x from $10 million in 2024 to $100 million in 2025. Latent Health, Tandem, Mandolin, Squad Health, Flow Auth, and others now auto-fill payer forms by pulling data from the EHR. Physicians submit claims instantly. Payers respond with their own AI like Optum Real to deliver instant coverage validation. We have an escalating AI arms race where both sides are investing billions to fight the same battle more efficiently. Neither side is asking whether the battle should exist.
CMS just brought AI-driven prior authorization to Traditional Medicare through the WISeR Model. They're paying AI contractors a percentage of denied claims to review services "prone to unnecessary use." We're importing Medicare Advantage's most criticized feature into a system that was already working well.

Dashevsky’s Dissection

So this captures everything I've learned about AI in healthcare over three years: we're treating symptoms, not root causes.

We're automating prior authorization instead of questioning why it exists, speeding up claims processing instead of redesigning how we pay for care, optimizing physician documentation instead of asking why physicians spend half their time on notes.

AI won't fix healthcare by making broken processes faster. It will only fix healthcare if it exposes the misaligned incentives we've been able to blame on complexity.

Right now, we can excuse inefficiency by saying the system is too complicated. But what happens when AI eliminates that excuse? When AI can automate prior auth in seconds, streamline care transitions seamlessly, and slash administrative costs? Suddenly the inefficiency becomes about incentive, not capability.

The healthcare organizations that thrive will be the ones willing to confront these misaligned incentives head-on and use AI to realign them around efficiency and patient care. The ones that don't will just run their broken systems faster.