Five years. That’s how long I spent trying to build an AI to diagnose mental health conditions. We raised money, hired a brilliant team, and built a model that was, on paper, incredibly accurate. And it failed. Completely.
I’m Sahin Boydas. You might know me from my last two companies, RemoteTeam and MovieLaLa, which were acquired by Gusto and Gfycat. Or maybe from my angel investments in companies like Anthropic, OpenAI, and Scale AI. I’ve been in Silicon Valley for a long time, and I’ve seen a lot of hype cycles. But I’ve never seen anything like the current frenzy around AI in healthcare.
Everyone is talking about how AI will revolutionize medicine. And it will. But not in the way most people think. I’m here to share the hard-won lessons from my own failed startup—the wins, the failures, and the counterintuitive strategies that actually work.
The Dream: An AI That Could See Depression
It started with a simple, powerful idea. What if we could diagnose depression and other mental health conditions just by analyzing a person’s voice and facial expressions? No more subjective questionnaires. No more waiting weeks for an appointment. Just a simple, objective, AI-powered diagnosis.
We called the company “Aura.” The plan was to build a mobile app that would record a short video of a patient, and our AI would analyze their speech patterns, tone of voice, and micro-expressions to screen for depression. We’d sell it to hospitals and insurance companies as a tool to identify at-risk patients and get them into treatment faster.
It seemed like a slam dunk. The tech was there. The market was huge. And we had a personal mission. I’ve seen friends and family struggle with mental health, and I knew that early detection could save lives.
Building the “Perfect” Model
We spent the first two years building our AI model. We hired PhDs in machine learning and computational linguistics. We licensed a massive dataset of clinical interviews and spent a fortune on data labeling. We fine-tuned our model for months, and the results were amazing. Our AI could detect signs of depression with over 90% accuracy in a lab setting.
We were on top of the world. We had a model that was more accurate than a human psychiatrist. We thought we had cracked the code. We were wrong.
The Brutal Truth
Here’s the part of the story that most founders don’t like to talk about. The part where the perfect model meets the messy reality of the real world.
Truth #1: A Great Model Is Not Enough
This is the single biggest lesson I learned. We were so focused on building a great AI that we forgot about everything else. We had a 90% accurate model that was completely useless in a clinical setting.
Why? Because we hadn’t thought about how it would actually be used. We hadn’t talked to doctors. We hadn’t thought about clinical workflows. We just assumed that if we built a great model, the rest would take care of itself.
It turns out that doctors don’t want a black box that spits out a diagnosis. They want a tool that helps them make better decisions. They want to see the data. They want to understand why the AI is making a particular recommendation. And they definitely don’t want to be replaced by an algorithm.
We also completely underestimated the regulatory hurdles. We thought we could just launch our app and start selling it. But in healthcare, you can’t just move fast and break things. You need to get FDA clearance, which is a long, expensive, and soul-crushing process.
Truth #2: The Data Problem Is Harder Than You Think
Everyone in AI talks about the importance of data. But in healthcare, it’s a whole different level of difficult. Getting high-quality, labeled data is a nightmare. The data is messy, it’s siloed in different hospital systems, and it’s protected by a mountain of privacy regulations.
We spent millions of dollars on a dataset that we thought was perfect. But when we started digging into it, we found all sorts of problems. The labels were inconsistent. The data was biased. And it didn’t reflect the diversity of the real world.
We ended up having to throw out half of our data and start from scratch. It was a painful and expensive lesson.
Truth #3: Doctors Are Not Your Users (At First)
We thought that if we could just convince doctors to use our app, we’d be golden. But we quickly learned that doctors are not the ones who make purchasing decisions in hospitals. It’s the administrators. And the administrators don’t care about how cool your AI is. They care about one thing: return on investment.
We had to completely change our sales pitch. Instead of talking about how our AI could improve patient outcomes, we had to talk about how it could save the hospital money. It was a tough pill to swallow, but it was the only way to get our foot in the door.
Truth #4: The Reimbursement Maze
This was the final nail in the coffin for Aura. We had a great model, we had a path to FDA clearance, and we had a sales pitch that was starting to resonate with hospital administrators. But we had no way to get paid.
In the US healthcare system, you don’t get paid for doing something cool. You get paid for doing something that has a billing code. And there was no billing code for an AI-powered mental health screening tool.
We spent months trying to navigate the reimbursement maze. We talked to insurance companies, we talked to Medicare, we even tried to lobby for a new billing code. But in the end, we just couldn’t make the numbers work.
What I'd Do Differently Today
So, what would I do differently if I were starting Aura today? A lot of things.
First, I would start with the user. I would spend the first six months talking to doctors, nurses, and hospital administrators to understand their pain points and how they actually work. I wouldn’t write a single line of code until I had a deep understanding of the clinical workflow.
Second, I would focus on a much narrower problem. Instead of trying to build a general-purpose diagnostic tool, I would focus on a specific use case where there is a clear and unmet need. For example, I might build a tool to help primary care physicians screen for postpartum depression.
Third, I would think about the business model from day one. I would have a clear plan for how I was going to get paid, and I would validate that plan with insurance companies and other payers before I started building anything.
The Future of AI in Healthcare
I’m still incredibly bullish on the potential of AI to transform healthcare. But I’m also a lot more realistic about the challenges. It’s not going to be a straight line. There are going to be a lot more failures like Aura along the way.
But that’s how progress happens. You try something, you fail, you learn, and you try again. And eventually, you build something that changes the world.
I’m not sure what my next company will be. But I know that it will be in the healthcare space. And I know that I’ll be a lot smarter this time around.
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.
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.
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.