Most founders think building a great AI model is enough. They're wrong. Here's the uncomfortable truth about what it really takes to succeed in the brutal, regulated world of healthcare AI.
I’ve seen it a hundred times. A brilliant team, a beautiful algorithm, a model that achieves near-perfect accuracy on a curated dataset. They raise a seed round, burn through it building the tech, and then slam headfirst into the brick wall of reality. The model that worked so well in the lab is useless in a real hospital. The data is a mess, the doctors won't use it, and the FDA just sent back a laundry list of questions they can't answer.
After two exits and over 200 angel investments in companies like Anthropic and Scale AI, I've developed a pretty good nose for what works. And what I'm seeing in AI for healthcare is a gold rush full of miners who have no idea how to dig. They're all chasing the glitter of the algorithm, but the real gold is buried deep in the messy, unglamorous work of implementation.
The 87% Illusion
Let's talk about that 87% number. It's real. My team and I spent months analyzing a dataset of over 10,000 clinical trials. We built a model that could predict the likelihood of a drug moving to the next trial phase with 87% accuracy. On paper, it's a phenomenal achievement. We could have slapped it on a landing page, raised a Series A, and called ourselves geniuses.
But I've been in this game too long. I knew that number was both a powerful asset and a dangerous illusion. An 87% prediction is worthless if you can't explain why the model made that prediction. It's worthless if it can't be integrated into the existing, chaotic workflow of a clinical research organization. And it's a massive liability if a regulator thinks you're just throwing black box predictions at them.
The real work wasn't getting to 87%. The real work was making that 87% useful.
Your Model is Not the Product
This is the single biggest mistake I see founders make. They fall in love with their model. They think the algorithm is the product. It's not.
The product is the entire system that makes the model's output actionable. It's the boring stuff. It's the data cleaning pipelines, the user interface that a tired, overworked clinician can actually understand, the compliance paperwork, and the customer support team that can explain a confidence interval to a skeptical doctor at 2 AM.
One of my portfolio companies, a promising AI diagnostics startup, learned this the hard way. They had a world-class model for detecting a rare cancer from medical images. But they spent all their money on data scientists and none on workflow integration. The doctors at their pilot hospital loved the idea of the tool, but nobody used it. It required them to log into a separate system, upload images manually, and wait for a result. It was a disruption, not a help. The company folded in 18 months.
Contrast that with another company I backed. Their model was, frankly, less impressive. But they spent a year just observing clinicians. They built a tool that plugged directly into the hospital's existing picture archiving and communication system (PACS). The AI's analysis showed up as a simple, color-coded overlay on the image the doctor was already looking at. It wasn't a new step in the process; it was an enhancement of the existing one. They are now in 50 hospitals and are about to close a major funding round.
Lessons from the Trenches
So, how do you avoid the traps? How do you build something that actually survives contact with the real world? Here are the hard-won lessons I've learned from my own startups and investments.
1. Solve a Workflow Problem, Not a Prediction Problem. Stop asking, "What can my AI predict?" Start asking, "What is the most painful, inefficient part of a doctor's or researcher's day?" Is it prior authorizations? Is it manually searching through patient records? Is it the mind-numbing process of screening candidates for a clinical trial? Find a workflow that is so broken, so universally hated, that a 10% improvement would be seen as a miracle. That's your entry point.
2. Embrace the Messy Data. Everyone wants a clean, perfectly labeled dataset. It doesn't exist in healthcare. You will be dealing with handwritten notes, conflicting EMR entries, and scans from a 20-year-old machine. Your ability to clean, structure, and make sense of this chaos is your single greatest competitive advantage. If you're only building for perfect data, you're building a fantasy. Invest in data engineers like your life depends on it. Because it does.
3. Build for Trust, Not Just Accuracy. In a regulated field, a black box is a non-starter. You need to build for interpretability from day one. This means:
- Explainability: Your model needs to be able to show its work. Why did it flag this patient as high-risk? What features in the data led to its conclusion?
- Confidence Scores: Never just output a binary "yes" or "no." Provide a confidence score. Let the human expert understand the model's certainty.
- User Overrides: The clinician must always have the final say. The AI is a co-pilot, not the pilot. Design your system so a doctor can easily review and, if necessary, override the AI's recommendation.
I once passed on an investment in a company with a brilliant algorithm because the founders were dismissive of explainability. Their attitude was, "The model is more accurate than a human, so the human should just trust it." That level of arrogance is a death sentence in healthcare.
The Real Future of AI in Healthcare
Forget the sci-fi headlines about robot surgeons. The real revolution in healthcare AI is happening in the background. It's in the automation of paperwork. It's in the optimization of clinical trial recruitment. It's in the simple tools that give a doctor back 10 minutes in their day.
Building in this space is not for the faint of heart. It's a long, brutal grind. It requires a deep respect for the complexity of the healthcare system and the expertise of the people working in it. But if you can solve a real, painful workflow problem and build a product that earns the trust of clinicians, you won't just build a successful company. You will fundamentally change how medicine is practiced.
Don't come to me with your fancy algorithm. Come to me when you've found a problem so agonizing that doctors are willing to try anything to solve it. That's when I'll write the check.
Frequently Asked Questions
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.
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.
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.