I get about five pitches a week for "AI in radiology." Every single one promises to "revolutionize medical imaging." Every single one has a beautiful deck showing a deep learning model highlighting a tumor with superhuman accuracy. And I turn down almost every single one.
It’s not because the tech is bad. Frankly, the models are getting incredibly good. It’s because the entire premise is flawed. The "AI for radiology" space has become a classic bubble, fueled by hype and a fundamental misunderstanding of how healthcare actually works. I’ve seen this movie before. I’ve lived it, both as a founder with two exits and as an angel investor in over 200 companies, including some of the biggest names in AI like Anthropic and OpenAI.
Most founders in this space are obsessed with building the perfect algorithm. They think the job is done once the model hits 99% accuracy. They’re wrong. Here's the uncomfortable truth about what it really takes to succeed in the brutal, regulated world of healthcare AI, and why the real opportunity isn’t where you think it is.
The Bubble: Why Your Radiology AI Startup Is Likely to Fail
I once invested in a promising startup that had a brilliant algorithm for detecting a rare type of fracture. The founders were PhDs from Stanford, the tech was solid, and they had a great dataset from a partner hospital. They burned through $5 million in two years and got exactly zero hospital contracts. They shut down last year. Why? They were so focused on the algorithm that they ignored the three things that actually kill healthcare startups: workflow, sales, and regulation.
The Data Moat is a Mirage
Every founder tells me they have a unique data advantage. They don’t. Hospital data is a disaster. It’s messy, unstructured, and locked away in ancient PACS systems that look like they were designed in the 90s. The data you get from one hospital won’t work with data from the hospital across the street because of different imaging protocols and machine vendors.
Building a model on a clean, curated dataset from a single academic partner is a great way to write a paper. It’s a terrible way to build a product. The real world is messy, and your model will break the second it sees data from a different scanner. The idea of a "data moat" in radiology is a fantasy used to impress VCs who don’t know any better.
The Workflow Integration Nightmare
A radiologist’s workflow is a chaotic symphony of clicks. They live in their PACS viewer, bouncing between studies, dictation software, and patient records. They are under immense pressure to read more studies, faster. The last thing they want is another screen, another login, another button to click.
Your "revolutionary" AI tool? To them, it’s just another interruption. If your solution isn’t seamlessly integrated into their existing software, it’s dead on arrival. It can’t just be accurate; it has to be invisible. It has to work so perfectly within their flow that they don’t even notice it’s there. This is a massive engineering and user experience challenge that most algorithm-obsessed founders completely underestimate. It requires deep partnerships with the legacy PACS vendors, which is its own special kind of hell.
The 24-Month Sales Cycle and Other Horrors
Let’s say you build a great product with seamless integration. Now you have to sell it. Selling to hospitals is a soul-crushing experience. The sales cycle is brutally long—I’m talking 18 to 24 months, minimum. You’ll have to convince the radiologist, the department chair, the IT department, the legal team, the purchasing committee, and the C-suite. Each one has their own budget, their own politics, and their own reasons to say no.
I’ve seen startups spend a year and a quarter of a million dollars just to get a single pilot project. And a pilot doesn’t guarantee a sale. Hospitals are littered with the zombie corpses of failed pilot projects. Unless you have a war chest to survive years of burning cash with zero revenue, you don’t stand a chance.
What to Focus on Instead: The Unsexy, High-Impact Opportunities
So if radiology is a bubble, where should ambitious founders focus? The biggest opportunities in healthcare AI aren’t in diagnostics. They’re in the "boring" stuff that makes the system run. It’s in the plumbing, not the facade.
1. Radical Healthcare Automation
For every dollar spent on a doctor, the US healthcare system spends another 50 cents on administrative overhead. It’s a bloated, inefficient mess of paperwork, billing codes, and phone calls. This is where AI can have the biggest and most immediate impact.
Think about automating the prior authorization process, which costs the system billions a year in wasted time. Or using AI to optimize hospital scheduling to reduce wait times and maximize resource use. Or AI-powered medical coding that eliminates human error and increases revenue.
These problems aren’t as sexy as finding cancer, but they are massive, expensive problems for hospitals. The ROI is immediate and easy to measure. When you can walk into a CFO’s office and say, "My software will save you $10 million next year by automating your billing," you will get their attention. The sales cycle is still long, but the value proposition is crystal clear.
2. The Final Frontier: AI in Drug Discovery
This is the high-risk, high-reward moonshot. Developing a new drug takes a decade and can cost over a billion dollars. Most of that time and money is spent on failed experiments. AI has the potential to fundamentally change the economics of drug discovery.
Companies are using AI to analyze genomic data, predict how molecules will behave, and design new proteins from scratch. This isn’t about fitting into a clinical workflow. It’s a pure R&D play. It’s about using massive computational power to solve one of the hardest scientific problems in the world. My investments in companies like Scale AI and Hugging Face have shown me the sheer power of large-scale models, and applying that power to biology is the next frontier.
It’s a long game, and the scientific risk is huge. But a single success could be worth billions and change the course of medicine. It’s a game for founders with deep technical expertise and the patience to pursue a 10-year vision.
3. The Human Connection: AI in Mental Health
The world is facing a mental health crisis. There aren’t enough therapists to meet the demand, and the stigma keeps many from seeking help. This is an area where AI can provide support at scale.
I’m not talking about AI replacing therapists. I’m talking about AI-powered tools that can provide personalized coaching, deliver cognitive behavioral therapy exercises, and offer a listening ear 24/7. The regulatory hurdles are lower than in diagnostics, and the need is immense. Building a trusted brand and a product that genuinely helps people is the key here. It’s less about the algorithm and more about empathy, design, and the human touch.
Stop Chasing the Bubble
The AI revolution in healthcare is real. It’s happening right now. But it’s not happening in the radiology reading room. It’s happening in the billing department, the research lab, and on the phones of people struggling with anxiety.
My advice to founders is simple: stop chasing the shiny object. Stop trying to build a better tumor detector. The world doesn’t need another one. Instead, go find a messy, expensive, unsexy problem that is costing the healthcare system billions of dollars. Go solve that. Be a plumber, not a painter. That’s where the real opportunity lies.
Frequently Asked Questions
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.
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.