What Can We Learn When Agents Fail?

In this week's episode of the Health Data Ethics Podcast, I dig into something that doesn’t get enough airtime in healthcare AI conversations: what we can learn from the projects that don’t work. A recent MIT study suggests that up to 95% of AI pilots don’t deliver a measurable return on investment. That statistic may feel discouraging, but it’s also a great jump off point for investigation. What's working, and what's not, and why? In this episode, I talk through two recent experiments with generative and agentic AI and try to pull out the lessons: Where agentic AI tends to break down in real-world workflows How governance and oversight need to evolve when AI systems are making decisions And why it may be more useful to think of AI as a new class of labor rather than a new type of software Healthcare has always been a hard place for innovation, not because we don’t want change, but because the consequences of failure are high, and the systems are complex. That makes it even more important to treat each failure, false start, or stalled pilot not as a sunk cost, but as feedback.