Dr. Pedro Teixeira: Thoughtful AI: The Honest Conversation Healthcare Isn't Having

Dr. Karen Litzy is joined by Dr. Pedro Teixeira, Vice President of AI Engineering at Prompt Health and former co-founder and CEO of Prediction Health, for a practical conversation about how AI fits into healthcare workflows. Pedro breaks down the differences among automation, AI, and agentic systems and explains why the best tools support clinicians rather than replace their judgment. This episode focuses on what clinic owners and clinicians should look for when evaluating AI tools: workflow fit, safety, latency, cost, and whether the system improves over time. If you want a grounded, non-hype conversation about AI in healthcare, this is a useful one.

Key topics

·       In this episode, Karen and Pedro compare thoughtful AI vs powerful AI and why workflow fit matters as much as model capability.

·       Pedro explains latency in practical terms, including why perceived delay matters and how feedback like loading states can make AI feel faster.

·       They unpack tokens, model size, and cost, including why input and output usage can change pricing so quickly.

·       Pedro shares how AI can work in the background across notes, codes, compliance checks, and analytics for clinic owners.

·       The conversation covers HIPAA, PHI, encryption, sandboxing, and limited tool access as essential safeguards for healthcare AI.

·       Pedro distinguishes automation, AI, and agentic systems, with examples of when rules are enough and when messy data really does call for AI.

·       They compare predictive vs generative models and explain why generating plausible text is not the same as forecasting a clinical outcome.

·       Karen and Pedro discuss how some clinics are using AI as a decision support tool, while still keeping humans in control of final decisions.

·       Pedro emphasizes that clinic owners should think in terms of systems, including what happens when AI works well and when it fails.

·       He closes with a simple evaluation question for any AI pitch: How does this system get better over time?