My inbox is a graveyard for 'AI in radiology' startups. Seriously. I get at least three pitches a week from founders convinced they’re going to revolutionize medical imaging. They all get the same answer: no.
It’s not because the tech isn’t clever. It’s because they’re all climbing the wrong mountain. They’re obsessed with building the perfect model, convinced that a 99.9% accuracy score is a golden ticket. They think they’ve built a better mousetrap and that hospitals will beat a path to their door. They’re about to get a very rude awakening.
I’ve been in the Silicon Valley trenches for over a decade, with two exits under my belt and over 200 angel investments in companies like Anthropic and OpenAI. I’ve seen this movie before. And I can tell you how it ends for most of these radiology AI companies: a slow, painful fade into irrelevance.
The Algorithm Is the Easy Part
Here’s the uncomfortable truth: building a high-performing diagnostic AI model in 2025 isn’t the massive achievement founders think it is. It’s table stakes.
With the power of foundation models and the incredible tooling from companies like Scale AI and Hugging Face, a small, sharp team can get to human-level performance on a narrow imaging task in a few months. Your model’s accuracy isn’t a competitive moat; it’s a feature. And it’s a feature your competitors can, and will, replicate in a heartbeat.
The real challenge isn’t the algorithm. It’s everything else.
I learned this the hard way. My first startup, MovieLaLa, had a brilliant movie recommendation engine. Technically, it was a marvel. But we poured all our energy into the tech and almost none into how it would actually fit into people’s lives. It was a painful lesson in humility. My second company, RemoteTeam, was the complete opposite. We obsessed over workflow. Gusto acquired us because we didn’t just build a tool; we built a system that integrated seamlessly into the messy, complicated reality of managing a remote team—payroll, compliance, and a hundred other non-technical headaches.
Healthcare is that challenge on steroids.
The Three-Headed Monster of Healthcare AI
When you try to sell an AI product to a hospital, you’re not just fighting other startups. You’re fighting a three-headed monster: workflow integration, data fragmentation, and a sales cycle from hell.
1. The Workflow Chasm
Radiologists don’t just sit in a dark room staring at images all day. They operate inside a complex, deeply entrenched workflow. It involves the Picture Archiving and Communication System (PACS), the Electronic Health Record (EHR), the Radiology Information System (RIS), and a dozen other pieces of ancient software that barely speak to each other. Your “revolutionary” AI tool? To a busy radiologist, it’s just another screen, another login, another interruption to a process that has been hammered into shape over decades.
Getting a radiologist to change their workflow is like trying to redirect a river with a teaspoon. You can’t just throw a new tool at them and hope it sticks. You have to integrate so deeply that your product feels like a natural extension of what they already do. That’s not an engineering problem. It’s a systems design, user experience, and change management problem of the highest order.
2. The Data Quagmire
Everyone loves to talk about the power of data in AI. The reality in healthcare is a nightmare of data fragmentation. Every hospital system has its own data formats, its own weird labeling conventions, and its own ironclad privacy protocols. The pristine, organized data you trained your model on in a lab looks nothing like the messy, inconsistent, and often incomplete data you’ll find in a real hospital.
Just getting access to high-quality, diverse data to start building a robust model is a monumental task. And once you’re deployed, you have to fight a constant battle against data drift. New machines, different patient populations, and changing clinical practices will cause your model’s performance to degrade. You need a massive, expensive infrastructure for monitoring, retraining, and redeploying. This is where the real, unglamorous work of AI begins.
3. The Sales Cycle from Hell
Selling to consumers is fast. Selling to small businesses is slower. Selling to hospitals can take years. You’re not selling to one person; you’re selling to a committee of doctors, IT staff, administrators, lawyers, and procurement officers. Each of them has their own agenda and a hundred reasons to say no.
I’ve watched promising startups with incredible technology burn through all their cash just trying to get that first hospital contract signed. The sales cycle is brutally long, incredibly expensive, and full of political landmines. You need a war chest and the patience of a saint. Most AI-first founders have neither.
Where the Real Money Is
So am I down on AI in healthcare? Absolutely not. It’s one of the areas I’m most excited about. But I’m not investing in the ten-thousandth company trying to read chest X-rays better.
I’m looking for the companies that are using AI to solve the unsexy problems. The boring, administrative, workflow-clogging problems that cause real pain for doctors and add billions in costs to the system.
Automating Clinical Documentation: Doctors are drowning in paperwork. They spend more time typing notes into an EHR than talking to patients. An AI that can listen to a conversation and automatically generate a clinical note is a billion-dollar opportunity. It doesn’t have to be perfect; it just has to be better than the soul-crushing status quo.
Simplifying Patient Communication: How many times have you left a doctor’s office completely confused about what to do next? AI can create personalized discharge instructions, answer common patient questions 24/7, and manage follow-up communication. This improves patient outcomes and frees up overworked nurses.
Fixing the Back Office: Think medical billing, insurance coding, and prior authorizations. This is the administrative plumbing of the healthcare system, and it’s a multi-billion dollar disaster. An AI that can reduce administrative overhead by even a few percentage points is a unicorn in the making.
These problems aren’t as glamorous as finding a tiny tumor on a CT scan. But they are real, they are painful, and they are where you can build a massive, sustainable business. Stop chasing the diagnostic bubble. Go find a real, messy, unsexy problem and solve it.
That’s how you’ll win. That’s where I’m placing my bets.
Frequently Asked Questions
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.