I once watched a founder with a brilliant AI model—one that could predict a rare form of cancer with 99% accuracy—completely fail to get a single doctor to use it. He’d spent millions on R&D, hired the best machine learning engineers, and had a technically superior product. But he made a classic mistake: he thought the tech was the hard part.
He was wrong. The hardest part of building clinical AI isn’t the algorithm; it’s earning the trust of the people who are supposed to use it.
After two exits and over 200 angel investments in companies like Anthropic and Scale AI, I’ve seen this pattern repeat itself. Founders, especially those from a pure tech background, march into healthcare with a solution-first mindset. They see doctors as end-users who will simply adopt a better tool. But medicine isn’t like other industries. A doctor’s workflow is a complex, high-stakes environment where a single error can have devastating consequences. Trust isn’t a feature you can add in the next sprint. It’s the entire foundation.
The Allure of the Perfect Algorithm
We all love a good story about a powerful AI. In diagnostics, we hear about models that can spot tumors in scans that are invisible to the human eye. In drug discovery, AI is sifting through billions of molecules to find promising new candidates. These are exciting developments, and I’ve invested in companies doing exactly that. But the hype often obscures a more mundane reality.
I remember a pitch from a startup that had developed an AI for reading chest X-rays. Their model was outperforming radiologists in a controlled environment. They had the data, the papers, the benchmarks. They thought they were ready for a Series A. I asked them a simple question: “How does a doctor use this in the middle of a chaotic ER shift?”
They didn’t have a good answer. Their "solution" involved the doctor stopping what they were doing, opening a separate application, uploading the X-ray, and waiting for the AI’s analysis. Anyone who has spent five minutes in an emergency room knows that’s a non-starter. The tool, as brilliant as it was, didn’t fit the workflow. It was a workflow interruption, not an enhancement.
That’s the first lesson: your AI is useless if it doesn’t seamlessly integrate into the existing clinical workflow. Doctors are not going to change their habits for your product unless the value is immediate and overwhelming. You have to meet them where they are.
Data: The Unstructured, Messy Reality
Silicon Valley runs on clean, structured data. We build models on neatly labeled datasets. In healthcare, that’s a fantasy. Clinical data is a mess. It’s fragmented across different systems, stored in different formats, and full of inconsistencies and errors. Electronic Health Records (EHRs) were designed for billing, not for research or AI.
One of my portfolio companies working on a predictive model for sepsis spent the first year of their existence not building an algorithm, but just trying to get usable data. They had to pull information from a dozen different sources—lab results, physician notes, vital signs monitors—and then spend months cleaning and standardizing it. The raw data was so noisy that their initial models were completely unreliable.
This is where a lot of founders get discouraged. They expect a plug-and-play data environment and instead find a digital hornet’s nest. But the companies that succeed are the ones that embrace this challenge. They understand that data acquisition and cleaning isn’t a preliminary step; it’s a core competency. They build teams that include not just data scientists, but clinical data managers and informatics specialists.
And then there’s the issue of privacy. HIPAA is not just a set of rules; it’s a fundamental design constraint. You can’t just pull a patient’s data and start training a model. You need to have robust systems for de-identification and security. This is another area where I see tech founders underestimate the complexity. They’re used to the relatively open data ecosystems of consumer tech. Healthcare is a walled garden, and you need to earn the right to be inside it.
The Black Box Is a Red Flag
Let’s say you’ve solved the workflow problem and you’ve managed to get your hands on some decent data. Now you have to convince a doctor to trust your AI’s recommendation. This is where the concept of "explainability" becomes critical.
Imagine you’re a doctor and an AI tells you a patient is at high risk for a heart attack. Your first question is going to be “Why?” You’re not going to put someone on a serious medication or recommend a procedure based on a black box algorithm’s say-so. You need to understand the reasoning behind the prediction. Is it because of their cholesterol levels? Their blood pressure? A specific combination of risk factors?
Many of the most powerful AI models, like deep neural networks, are notoriously difficult to interpret. They can find patterns that humans can’t, but they can’t always explain how they got there. In a field like medicine, that’s a major problem. A doctor is ultimately responsible for their decisions. They can’t delegate that responsibility to an algorithm they don’t understand.
This doesn’t mean you can’t use complex models. But it does mean you need to invest in explainable AI (XAI) techniques. You need to be able to show the doctor the key features that drove the AI’s recommendation. This could be a heatmap on an image highlighting a suspicious area, or a list of the top contributing factors to a risk score. The goal is not to replace the doctor’s judgment, but to augment it. The AI provides a new piece of evidence, and the doctor makes the final call.
Building Trust, One Doctor at a Time
So how do you actually build that trust? It’s not about marketing or sales pitches. It’s about collaboration. It’s a slow, painstaking process of working with doctors, listening to their feedback, and iterating on your product.
One of the most successful clinical AI companies I’ve invested in started by partnering with a single hospital. They didn’t try to sell their product to a thousand doctors at once. They found a small group of early adopters who were excited about the technology and willing to work with them. They spent months in the hospital, shadowing doctors, and understanding their pain points. They treated the doctors as design partners, not as customers.
Their initial product was clunky and had a lot of problems. But because they had built a strong relationship with their early users, they got honest, constructive feedback. They iterated quickly, releasing new versions of the product every few weeks. Over time, they built a tool that the doctors loved because it was their tool. They had helped create it.
Once they had a successful pilot at one hospital, they had a powerful case study. They could go to other hospitals and say, “Look, this is what we did at Hospital X, and here are the results.” They had proof that their AI could work in a real-world clinical setting. That’s far more valuable than any benchmark on a leaderboard.
The Real Opportunity
Building a successful clinical AI company is one of the hardest things you can do in tech. The challenges are immense: the regulatory hurdles, the data problems, the long sales cycles, the deeply entrenched workflows. But it’s also one of the most meaningful.
We are at the very beginning of a revolution in medicine. AI has the potential to make healthcare more proactive, more personalized, and more effective. It can help doctors diagnose diseases earlier, develop new treatments, and deliver better care to more people. But that potential will only be realized if we build tools that doctors trust and use.
My advice to founders in this space is this: fall in love with the problem, not your solution. Spend more time in hospitals than you do in front of a computer. And remember that the person you’re building for isn’t a user; they’re a partner. If you can do that, you won’t just build a great product. You’ll build a company that actually makes a difference.
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.
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.