Why Smarter AI Agents Still Break With Juhi Parekh (GM of Key Frontier AGI Accounts at Turing)

In this episode of AI Explained, we are joined by Gary Kotovets, Chief Data & Analytics Officer at Dun & Bradstreet, where he leads data, analytics, and AI strategy across the company's commercial graph of 650-plus million businesses. Before this role, he spent nearly two decades at Bloomberg as global head of data acquisition and management, building the rigor around data quality that now shapes how he thinks about agentic AI. Gary breaks down what it takes to make enterprise data agent-ready — lineage and provenance, roughly 100 billion data quality checks run against source data, and a strict internal policy that every AI-generated answer must be able to show where it came from. He and Krishna dig into why hallucination is a model problem rather than a data problem at D&B, how a shared tools library with built-in rules keeps agents from drifting off script as workflows move from single-turn chat to multi-step autonomous tasks, and where governance, small language models, and agent-to-agent interactions are headed over the next few years. They close with a rapid-fire round covering data versus models, RAG versus fine-tuning, and the most overhyped and underrated ideas in enterprise AI right now.

In this episode of AI Explained, we are joined by Maryam Ashoori, PhD, VP of Product and Engineering for watsonx.governance at IBM, where she leads the teams building IBM's platform for governing AI models and agents across the enterprise. Before this role she headed product for watsonx.ai, led engineering for Lyft's bikes and scooters operations, and spent six years at IBM Research working on emerging technologies including AI and quantum computing. Maryam breaks governance down into three foundations — visibility, control, and accountability — and explains why enterprises can only govern the AI they can see while shadow AI keeps agents and models out of view. She and Krishna dig into what an AI control plane should actually do (define, implement, enforce, and track controls), why accountability is the top challenge enterprises cite as agent adoption scales, and how third-party risk, business continuity, and an evolving regulatory landscape are reshaping what "in control" means. They close with a rapid-fire round covering copilots vs. autonomous agents, frontier vs. small models, and the one AI belief Maryam has changed her mind about.

In this episode of AI Explained, we are joined by Juhi Parekh, GM of Key Frontier AGI Accounts at Turing. Juhi brings experience across the full AI stack, from applied AI and foundation models to data infrastructure, with prior product roles at Apple, Amazon, Niantic, Spatial, and Samsung Research US, where she focused on commercializing frontier AI. She explains how Frontier Labs curates hard datasets that maximize information gain rather than raw difficulty, why the sweet spot for reinforcement learning tasks is problems frontier models fail at least 30 percent of the time, and how long-horizon, real-world workflows are pushing agents to take on more complex work. She also shares the usual suspects when agents break in production (inaccurate tool calls, consistency gaps, permissioning, and output format), why training a capable model and building a reliable agent are two different problems, and why the winners will be the organizations that safely expand agent freedom as guardrails improve.

In this episode of AI Explained, we are joined by Jeff Malnick, VP of Engineering for Developer and AI Products at 1Password. Jeff brings deep expertise in distributed systems, secrets management, and security infrastructure from his work at 1Password and previously HashiCorp, with a focus on making the secure path the default for developers and now AI agents.  He explains why agent identity has suddenly become urgent: AI agents are machine workloads with reasoning capabilities, which breaks the assumption behind traditional OAuth flows where permissions only need to be granted once. He walks through 1Password's framework of three agent identity models (delegated, bounded, and fully autonomous), why the laptop is the hardest environment to secure now that file system access effectively hands an unlocked machine to a reasoning stranger, and why bearer tokens and passwords are the wrong primitive for agents. He also shares how policy decision and enforcement points need to evolve to pull human intent through to just-in-time authorization, why credentials should never enter LLM context, and where federated versus distributed identity is heading over the next three to five years.

In this episode of AI Explained, we are joined by Stacey Johnston, MD, Chief Information and Digital Execution Officer at Beacon Health System. Dr. Johnston brings a rare dual perspective to healthcare AI. She trained as a physician and hospitalist before moving into health informatics, giving her firsthand insight into where technology helps clinicians and where it gets in the way. She discusses how Beacon built its AI governance council from scratch, the policies and vendor risk assessments that gate every new tool, and why requiring a defined ROI before approval has become a forcing function for disciplined adoption. She also shares how agentic AI is already delivering results in scheduling, autonomous benefits verification, and colon cancer screening, why ambient listening drove $10,000 in additional per-physician revenue over 12 months, how clinician trust is earned incrementally through seamless workflow fit and real time savings, and what a federated but centrally monitored AI model could look like as health systems scale.

In this episode of AI Explained, we are joined by Jeff Dalton, Head of AI and Chief Scientist at Valence. Jeff has spent two decades at the intersection of research and industry, from building early conversational search benchmarks at Carnegie Mellon and Microsoft to leading the AI behind Nadia, Valence's purpose-built enterprise coaching assistant. He discusses the fundamentals of agentic system design that still hold from classical AI theory, why evaluation has to come before the prompt, how he approaches memory as a first-class object in coaching systems, and the defense-in-depth approach to guardrails that keeps complex agents safe across diverse enterprise deployments.

In this episode of AI Explained, we are joined by David Kenny, Executive Chairman of the Board of Nielsen and veteran AI leader.  He discusses how to prevent AI agents from going rogue, sharing insights from Nielsen's "Ask Nielsen" platform and emphasizing the importance of compound AI systems, real-time control planes, "generally accepted trust principles" (GATP) for AI, fit-for-purpose models, production cost management, and change management in moving from pilot to production.

In this episode of AI Explained, we are joined by Gary Stafford, Principal Solutions Architect at AWS Strands Agents. He delves into how enterprises choose between AI/ML and agentic approaches, patterns for multi-agent systems, and the role of MCP. Gary also shares real-world use cases and practical guidance on safety, scaling, and delivering enterprise-ready agent systems.

In this episode of AI Explained, we are joined by Jayeeta Putatunda, Director of AI Center of Excellence at Fitch Group.  She discusses essential lessons learned from building and deploying AI agent systems, including challenges in moving from concept to production, key evaluation metrics, and the importance of observability and guardrails in ensuring reliable AI systems.

In this episode of AI Explained, we are joined by Nate B. Jones, AI strategist. He explores high-level advice for organizations, technical ideas such as prompting and application architecture, and the current state of agent adoption. Key topics include challenges in building production-ready agents, architectural decisions, and ensuring ROI from these agents.