Show Notes
In 2024, a single cyber attack exposed the medical records of 190 million Americans. As healthcare organizations rush to adopt AI—with 38% now using it regularly—a new crisis is emerging: how do we harness AI's transformative power while protecting the most sensitive data we possess? This episode tackles the critical intersection of AI innovation and healthcare data security, where the stakes couldn't be higher. Sam and Mac reveal alarming statistics that healthcare executives can't afford to ignore: AI privacy incidents surged 56.4% in 2024, with 72% of healthcare organizations citing data privacy as their top AI risk. The average healthcare breach now costs $11.07 million per incident, yet only 17% of organizations have technical controls in place to prevent data leaks. The math is terrifying—and the problem is accelerating. The conversation explores how AI fundamentally changes the threat model in healthcare. Unlike traditional software that processes data according to fixed rules, AI models can unintentionally retain sensitive patient information from training data, creating new vulnerabilities that standard security practices weren't designed to address. Shadow AI—unauthorized AI tools used by employees handling sensitive data—poses massive compliance risks that most organizations haven't even begun to map. But this isn't just a doom-and-gloom episode. Sam and Mac outline emerging solutions that could reshape how healthcare handles AI and data security. Federated learning allows AI models to train across multiple institutions without patient data ever leaving its original location, enabling collaboration without exposure. Synthetic data can mimic real patient populations for AI training without using actual patient information, dramatically reducing privacy risks while maintaining analytical value. Looking forward, the episode emphasizes that stronger regulations and compliance practices aren't obstacles to AI adoption—they're prerequisites for sustainable innovation. Patient trust is healthcare's most valuable asset, and once lost through a major AI-related breach, it may be impossible to recover. The organizations that will thrive in the AI era are those that treat data protection not as a compliance checkbox but as a competitive advantage and moral imperative. Key topics covered: • The 2024 cyber attack exposing 190 million American medical records • Why 72% of healthcare organizations cite data privacy as their top AI risk • The 56.4% surge in AI privacy incidents involving PII (personally identifiable information) • Healthcare breach costs: $11.07 million average per incident • Shadow AI risks: unauthorized tools handling sensitive patient data • Why only 17% of organizations have adequate technical controls • How AI models unintentionally retain sensitive training data • Federated learning: training AI without data leaving institutions • Synthetic data: mimicking real populations without using actual patient information • The regulatory landscape and need for stronger compliance frameworks • Balancing innovation velocity with responsible AI practices • Privacy-preserving techniques: differential privacy and secure multi-party computation • Patient trust as healthcare's most critical asset in the AI era • Practical governance frameworks for healthcare AI implementation This episode is essential listening for healthcare executives navigating AI adoption, data security professionals protecting sensitive information, technology leaders implementing AI systems, and anyone concerned about the privacy implications of AI in medicine. Sam and Mac cut through the hype to deliver actionable insights on one of healthcare's most pressing challenges: how to innovate responsibly in an era where a single breach can expose hundreds of millions of records.
More from Inside AsembleAI: DeepTech, AI & Science
Vanessa Larco has sat on both sides of the table. She built products at Microsoft, Disney, Twilio, and Box, then led investments at NEA, including Mejuri and Kindred, and watched Robinhood scale to its IPO as a board observer. Sam and Mac talk with her about what actually holds up once the pitch deck closes. Topics covered: Why a founder’s roadmap, and what’s missing from it, shows how they make decisions Skill, luck, and the dynamics no investor can predict Product market pull: users who would be upset if you took the product away Why enterprise AI gets the capital, and where consumer AI is still open The hybrid go-to-market challenge of selling AI to small businesses AI-native companies vs. companies that use AI behind the scenes without selling it The “why now” framework behind Mejuri and Kindred Lessons from Robinhood’s founders on urgency, obsession, and speed Agent traffic vs. human traffic, and why product teams will need agent metrics How a venture firm uses AI agents, and why conviction in diligence stays human Guest Bio: Vanessa Larco is co-founder of Premise Venture Partners and was previously a Partner at NEA, a fund with over $25 billion in AUM, where she led investments in brands like Mejuri and Kindred. Before venture, she built products at Microsoft, Disney, Twilio, and Box. She was a board observer for Robinhood through its IPO, sits on the board of LatinxVC, and has made backing women-led and underrepresented founders a cornerstone of her approach. Connect: Find Vanessa on LinkedIn, where she says she’s most responsive. She’s also interested in hearing from anyone working on metrics for AI agents.
Rowan Stewart has watched companies rush AI products to market without the governance to back them up, first at BCG and now at Transcend, where the clients include Robinhood, Brex, and Groupon. Sam and Rowan discuss why governance works best as infrastructure from day one, and why agents raise the stakes. Topics covered: "Pilot purgatory": why promising AI proofs of concept die at the edge of production The cost of rushing: lost customer trust and bad press matter more than regulatory fines Embedded vs. bolted-on governance: a stack of paperwork and a handshake vs. encoded, auditable controls Deterministic controls for non-deterministic systems, and why a system prompt isn't a real control Human in the loop at scale: a 100-person company with five agents per employee, and how SMBs should adapt Why agents shouldn't inherit their owner's permissions, and why MCP tool restrictions aren't sufficient Agent observability: knowing what your agents are doing across the enterprise How guardrails let teams move faster, and using signals like denial rates to tune policy AI governance vs. responsible AI: business outcomes vs. net benefit to humanity The one gap to fix first: do you know what agents you even have running? Aligning incentives between legal/compliance and product/data science teams Guest Bio: Rowan Stewart is an AI and Data Safety Product Leader at Transcend, which helps companies like Robinhood, Brex, and Groupon govern and activate their data responsibly. Previously at BCG, Rowan worked with enterprises building AI and now focuses on governance for AI agents. Connect: Find Rowan Stewart on LinkedIn. Rowan's profile link is in the show notes.
In this episode of Inside AsembleAI, host Mac Goswami sits down with Dr. Latha Karthigaa — Director & Head of AI Governance at the Global AI Certification Council (GAICC) — on why most companies that think their AI is governed are dangerously wrong, and what responsible AI actually looks like. What's Covered: "A Policy Is Not Governance" — The biggest myth in enterprise AI. Latha's highway analogy: governance is the guardrail that lets you drive faster, not the thing slowing you down. 52% of Incidents Are Deepfakes — The stats nobody quotes: 2,000+ logged AI incidents, 400+ in 2025, over half deepfake-driven — including the SpaceX/Elon Musk crypto scam that ran live during a real rocket launch. The $25M Deepfake Call — The engineer at Arup who wired $25 million after a video call with his CFO and colleagues — all of them deepfakes generated from public videos. Why employee training, not just policy, is the real defense. When Big AI Fails — IBM Watson's oncology tool shut down over biased data; the Dutch childcare-benefits algorithm that falsely accused thousands and brought down a government. Why Latha says the best governance pros are empathetic humans first. Who Owns the Risk? — "ChatGPT isn't responsible — you are." Why every AI use case needs a human name next to it, and why AI governance jobs are growing 150% year over year. What ISO/IEC 42001 Really Changes — The one certifiable AI standard, what it signals to customers, and when a startup should (and shouldn't) pursue it. Key Quote: "Policy is like potential power. Execution is the real power." Connect with Latha: LinkedIn: https://www.linkedin.com/in/lathakarthigaa/ GAICC: https://www.gaicc.org/ · YouTube: Latha – AI Governance Follow and subscribe to AsembleAI: Spotify | Apple Podcasts | Amazon Music | iHeart Radio | YouTube | Substack
This episode was recorded live from the Ai4 conference podcast pavilion, where host Mac Goswami sat down with Barr Moses — CEO & co-founder of Monte Carlo, creator of the data observability category and now builder of the agent trust platform — about what it actually takes to make AI trustworthy in production. What's Covered: "AI Is Wrong in Your Face" — Barr's framing of the trust gap: dashboards were always wrong quietly in the background; AI is wrong out loud, and it'll argue with you. Why trust is the biggest thing standing between pilots and production. The Four Layers of Agent Failure — Context, performance, behavior, and output. Why all four can look perfect and the agent still fails — and why you have to watch all of them together. The Flight That Already Left — The airline agent that recommended a flight that departed that morning. The agent was fine; the context was stale. The most surprising failure mode nobody plans for. Where to Start — Make ONE agent great, not a hundred. And why the hardest first step is simply defining what "good" even looks like. The Reinforcement Loop — The idea Barr's most excited about: agents that self-identify what went wrong, propose a fix, submit a PR for human approval, and use it as tomorrow's baseline. Agents that rebuild themselves every day — running in production today. 100% AI-First — Why every line of Monte Carlo's code is AI-generated, how it made them 3–5x faster, and Barr's stoplight analogy for where human-in-the-loop is heading. Key Quote: "Dashboards are wrong in the background. AI is wrong in your face — it'll argue with you." Connect with Barr: LinkedIn: Barr Moses : https://www.linkedin.com/in/barrmoses/ Monte Carlo: https://www.montecarlo.ai Follow and subscribe to AsembleAI: Spotify | Apple Podcasts | Amazon Music | iHeart Radio | YouTube | Substack
ActualyzeAI came out of stealth just days before this conversation. Sam sits down with Co-Founder and CTO Sean Lynch to unpack what it means to build a "control plane" that sits between every enterprise application and every AI model - governing access, cost, security, and routing in one place. Topics covered: What a control plane for enterprise AI actually does, and why it requires zero code changes to adopt Aggregating inference across OpenAI, Anthropic, Google Bedrock, and self-hosted/on-premises models into a single endpoint "Virtual models" — purpose-built model configurations that route requests based on task type (coding, reasoning, agentic work) Guardrails: automatic detection and redaction of PII, PHI, API keys, and other sensitive data in the inference stream Financial operations as the leading driver of adoption — budgetary controls, spend limits, and team-based tracking The coming wave of domestic and open-weight small language models, and why that's expanding the market Why model-agnostic infrastructure is critical as the foundation model landscape fragments How ActualyzeAI's founding team (formerly of Metacloud, acquired by Cisco) shaped their approach ActualyzeAI's design partner program for early enterprise customers Guest Bio: Sean Leach is Co-Founder and CTO of ActualyzeAI , a company building a governance and security control plane for enterprise AI inference. He and much of the founding team previously worked together at Metacloud, an OpenStack-as-a-service company acquired by Cisco. Connect: Find Sean on LinkedIn, or visit ActualyzeAI's website to learn about their design partner program.