Your brilliant AI model is practically worthless.
There, I said it. Harsh? Maybe. But after a decade in the Silicon Valley trenches, two exits, and writing checks to over 200 startups, including some of the biggest names in AI, I’ve earned the right to be blunt. I see founder after founder make the same critical mistake, especially in healthcare. They spend millions and years of their lives perfecting a model in a lab, convinced that technical elegance will conquer the world.
They’re wrong. Dead wrong.
I’m going to tell you the story of what it really takes. Not the sanitized version you read on tech blogs, but the messy, unfiltered, behind-the-scenes battle of bringing an AI-powered medical device from a Jupyter notebook to a hospital. This is the story I wish someone had told me before I dove in.
The Great Delusion: My Model is 99% Accurate!
We started with a killer idea in AI-powered radiology. The concept was simple: use a deep learning model to detect early signs of a specific, aggressive cancer from standard MRI scans. Our team was stacked with PhDs from Stanford and Berkeley. We raised a solid seed round, and within a year, we had a model that was, in a lab environment, achieving 99.3% accuracy. We thought we’d cracked it. We were already mentally planning the Series A party.
This is the peak of what I call the “Founder’s Delusion.” You’re so in love with your technical achievement that you can’t see the massive iceberg dead ahead. The model is the first 1% of the journey. The other 99% is a brutal slog through data, regulations, and hospital politics.
The Data Moat is a Myth (At First)
Everyone talks about building a “data moat.” The reality is you start with a data puddle. We needed thousands of diverse, annotated scans to make our model robust. Where do you get them? Hospitals. And hospitals guard their data like Fort Knox. It took us six months of non-stop meetings, legal reviews, and begging just to get our first pilot agreement with a single hospital system.
And the data we got? It was a disaster. Scans from different machines, in different formats, with missing patient information, and radiologist notes that were more like cryptic poems than structured data. Our top-tier data scientists spent 80% of their time just cleaning and labeling data. It was expensive, mind-numbing work. The glamorous AI future we pitched to investors felt very far away.
Welcome to Regulatory Hell
If you think building the model was hard, you haven’t met the FDA. Getting a medical device approved is a multi-year, multi-million dollar ordeal designed to break your spirit. We were pursuing a 510(k) clearance, which is supposedly the “faster” path. It was anything but.
We had to document every single line of code, every decision, every parameter tweak. We had to prove our model’s performance on a statistically significant, independent dataset. The submission document was over 1,000 pages long. We hired three consultants and a full-time regulatory affairs manager just to handle the paperwork.
Six months after our submission, we got a letter back from the FDA. It was a list of 87 questions and requests for additional information. One of their concerns was that our model’s performance varied slightly between scans from GE and Siemens machines. It was a tiny, statistically insignificant difference, but it sent us back to the drawing board for another four months of data collection and model retraining. That one question cost us nearly half a million dollars in salaries and cloud computing bills.
The Last Mile: Where AI Meets Reality
Let’s say you survive the data and regulatory nightmare. You have your clearance. Now you have to actually sell this thing. And you quickly discover the most painful truth of all: the best AI is useless if it doesn’t fit into a doctor’s workflow.
Radiologists are some of the busiest people on the planet. They live inside their Picture Archiving and Communication System (PACS), a piece of software that often looks like it was designed in 1998. They don’t have time to open a separate application, log in, and upload a scan to get our AI’s opinion. Our initial design did exactly that. The feedback from our first pilot users was brutal. “I don’t have time for this,” one of them told me. “It’s a cool tool, but it makes my job harder, not easier.”
That comment hit me like a punch to the gut. We had spent years building this incredible technology, and it was being defeated by a bad user interface. We had to scrap our standalone app and spend another nine months building a complex, painful integration directly into the leading PACS systems. It was a technical nightmare of legacy protocols and undocumented APIs, but it was the only way.
The Real Playbook
We eventually got there. We got our integrations working, we closed our first few hospital contracts (after an average 18-month sales cycle), and the company was eventually acquired. But the journey was nothing like I expected.
If I were to do it all again, I’d ignore the conventional wisdom. Here’s what I’d do:
- Start with the workflow, not the model. Shadow a doctor for a month. Find the single most annoying, time-consuming part of their day. Build a tool for that. Even if it’s not a sexy AI problem.
- Become a service company first. Before you write a single line of AI code, sell a service where a human expert performs the task you want to automate. Use this to get real-world data, understand the workflow, and build relationships with customers. You get paid to do your R&D.
- Talk to the FDA on day one. Don’t wait until you have a finished product. Engage them early and often. They are not the enemy. They are your partners in navigating a complex process. It will save you millions in the long run.
Building an AI company in healthcare is not for the faint of heart. It’s a messy, brutal, and often irrational journey. It will test your patience and your sanity. But if you can survive the gauntlet, if you can focus on the real-world problems beyond the elegance of your model, you can build something that actually matters. Something that saves lives. And that’s a story worth telling.
Frequently Asked Questions
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.
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.