AI in Healthcare: What’s Actually Working and What’s Still Hype
Every year since 2020, someone declares it “the year AI transforms healthcare.” And every year, the reality is more nuanced than the headlines suggest. But 2026 is genuinely different — not because of some magical breakthrough, but because the boring stuff finally started working.
Diagnostics: Where AI Is Legitimately Saving Lives
Let’s start with what’s actually working, because there’s real progress worth talking about.
AI diagnostic tools are now deployed in hundreds of hospitals worldwide, and the results are hard to argue with:
Medical imaging. Companies like Zebra Medical Vision and Aidoc have AI systems reading chest X-rays, mammograms, and retinal scans with accuracy that matches or beats specialist physicians. Not in lab conditions — in actual clinical settings, processing real patient data.
The key stat: AI systems are now detecting cancers, strokes, and heart disease before symptoms appear, with over 85% diagnostic accuracy. That’s not replacing radiologists — it’s giving them a second pair of eyes that never gets tired and never misses a shift.
Pathology. AI-powered pathology is catching things human pathologists miss. Paige AI got FDA clearance for their prostate cancer detection system, and it’s finding cancer in biopsies that were initially read as negative. Think about what that means for patients who would’ve been told “you’re fine” and sent home.
Retinal screening. This is probably the most mature AI healthcare application. Diabetic retinopathy screening with AI is now standard in many countries. Patients get screened at their regular doctor’s office instead of waiting months for a specialist appointment.
Drug Discovery: Faster, But Not Magic
The drug discovery hype has been intense, and I want to be honest about where things actually stand.
AI is genuinely accelerating the early stages of drug discovery. Machine learning models can screen millions of molecular compounds in days instead of months. They can predict protein structures (thanks to AlphaFold and its successors) and identify promising drug candidates faster than traditional methods.
But here’s the reality check: faster discovery doesn’t mean faster drugs. Clinical trials still take years. Regulatory approval still takes years. The bottleneck was never “we can’t find promising molecules fast enough” — it’s everything that comes after.
What AI is doing well in 2026:
- Identifying drug candidates 60-70% faster than traditional screening
- Optimizing clinical trial design (better patient selection, adaptive protocols)
- Predicting drug interactions and side effects before trials begin
- Repurposing existing drugs for new conditions
What AI isn’t doing: replacing the fundamental biology of testing drugs in humans. That part is still slow, expensive, and necessary.
The Agentic Healthcare Shift
Here’s the 2026 development that I think is underrated: agentic AI is entering healthcare workflows.
Not as a diagnostic tool — as an operational backbone. AI agents are now handling:
Administrative tasks. Scheduling, insurance pre-authorization, medical coding, referral management. These are the tasks that burn out healthcare workers and delay patient care. AI agents are handling them faster and more accurately than the manual processes they replace.
Clinical documentation. AI scribes that listen to doctor-patient conversations and generate clinical notes in real-time. Doctors I’ve talked to say this alone saves them 1-2 hours per day. That’s 1-2 more hours of actually seeing patients.
Care coordination. AI agents that track patient follow-ups, flag missed appointments, and coordinate between specialists. The boring logistics that fall through the cracks in busy hospitals.
What’s Still Broken
I’d be doing you a disservice if I didn’t talk about the problems:
Data silos. Hospital systems still don’t talk to each other. Your medical records at Hospital A might as well not exist when you show up at Hospital B. AI can’t fix healthcare if it can’t access the data.
Bias. AI diagnostic tools trained primarily on data from one demographic perform worse on others. This isn’t theoretical — studies have shown AI skin cancer detectors that work great on lighter skin and poorly on darker skin. The training data problem is real and not fully solved.
Regulation lag. The FDA’s approval process for AI medical devices is getting faster, but it’s still not keeping pace with the technology. By the time an AI tool gets cleared, the model it’s based on might be two generations old.
Trust. Many doctors still don’t trust AI recommendations, and honestly, that’s not entirely unreasonable. “The AI said so” isn’t a diagnosis. Building trust requires transparency about how AI reaches its conclusions, and most current systems are still black boxes.
Where This Goes Next
My prediction for the rest of 2026: the biggest impact won’t come from flashy new AI capabilities. It’ll come from better integration of existing AI tools into clinical workflows.
The hospitals that figure out how to make AI a smooth part of their operations — not a separate system doctors have to learn — will see the biggest improvements in patient outcomes and staff satisfaction.
The technology is ready. The implementation is what needs work. And that’s actually good news, because implementation problems are solvable. We just need to stop chasing the next breakthrough and start making the current tools work better.
🕒 Last updated: · Originally published: March 12, 2026