NVIDIA AI Podcast Episode 261 Summary

Title: CVS Health and Aible are Delivering Enterprise AI with Rapid Prototyping, Agents, and Reasoning Models
Host: Noah Kravitz
Guests: Tony Ambrosi (SVP & Chief Digital and Technology Officer, Pharmacy & Consumer Wellness, CVS Health), Arijit Sengupta (Founder & CEO, Aible)
Date: June 18, 2025

Episode Overview

This episode explores how CVS Health and Aible are building and deploying enterprise-scale AI solutions by harnessing rapid prototyping, agentic AI frameworks, and advanced reasoning models. The conversation dives into the evolving AI landscape, how businesses can implement AI responsibly, and the increasing intelligence and autonomy of AI agents. Practical considerations, philosophical perspectives, and future-forward thinking are woven throughout, with both guests offering expert insights based on years of collaboration.

Key Discussion Points & Insights

1. Guest Backgrounds & The Foundation of Their Partnership

• Tony Ambrosi recounts a career at the intersection of technology and healthcare, emphasizing his passion for leveraging innovation to better serve customers and employees. He and Arijit first collaborated at Disney in the mid-2010s, sparking a mutual interest in applying machine learning for tangible improvements.

  • “[Machine learning] can transform a lot of… make a lot of things better.” – Tony Ambrosi [02:03]

• Arijit Sengupta brings a background in augmented analytics (BeyondCore, acquired by Salesforce as Einstein Discovery). His book, "AI is a Waste of Money," highlights pitfalls in enterprise AI—especially the disconnect between business users and data scientists.

  • “One of the most common reasons [AI projects] fail was disconnect between the business user and data scientists… it doesn't matter what you build unless the business stake adopted.” – Arijit Sengupta [02:34]

2. Rapid Prototyping and Implementation at Scale

• Prototyping Process:

  • CVS and Aible operate on a 48-hour rapid prototype standard; projects must present a working prototype to business stakeholders for feedback within two days.
  • Full-scale, impactful AI solutions are expected within 30 days after initial approval.
  • Key Reason: Fast cycles keep projects relevant and ensure alignment with business needs, avoiding “the graveyard of things that never go anywhere.” – Tony Ambrosi [03:20]

• Feedback and Iteration:

  • Immediate exposure to business users is essential: “If business is not involved, doesn’t matter.” – Arijit Sengupta [04:51]

Memorable Moment:

• “You talked for a week about what we want to do, then we go and do it for two weeks… then come back for another week and evaluate… that, and then so you have this cycle that takes months. Everybody will forget about what they wanted to do in the first place.” – Tony Ambrosi [03:20]

3. Platform Architecture: Aible and NVIDIA

• Aible’s Role:

  • Provides a low-code/no-code agent development platform that abstracts technical complexity.
  • Uses NVIDIA’s NIM, NeMo, and DGX cloud infrastructure.
  • Allows CVS teams to build specialized AI applications, focusing on outcomes rather than engineering.
    • “Tony’s team doesn’t have to worry about that [infrastructure].” – Arijit Sengupta [06:38]

• Responsible AI:

  • Critical in healthcare and consumer settings. Applications must keep the human in the loop, prioritize safety, and focus on responsible innovation.
  • “We have a particular focus on what we call responsible AI… how do we help our colleagues with things that probably AI can do in terms of automation while also, always keeping the human in the loop and in control.” – Tony Ambrosi [08:21]

4. Practical AI Applications at CVS (Non-Confidential)

• Focus is on supporting customer experience and internal automation (e.g., “summarizing invoices,” “procedural queries like ‘When is my prescription ready?’”).
• Conversational AI projects are being implemented for real-time customer support, especially for pharmacy and retail queries—without replacing pharmacists’ roles in critical tasks.

5. The Rise of AI Agents and “Agentic” AI

• Agent Definition and Role:

  • AI agents combine intelligence (via LLMs) and action (process automation), bridging the gap between insight and tangible work.
    • “Agents came at the intersection… of intelligent machines through large language models and… process automation.” – Tony Ambrosi [10:43]

• Levels of Autonomy:

  • Researchers describe a scale from 0 (manual) to 5 (fully autonomous, no human in the loop). Most enterprise use cases are at levels 2-3, with some automation and human oversight.
    • “Most of us are somewhere in the 2 to 3…” – Tony Ambrosi [13:29]

• Agent Capabilities:

  • Agents can now manipulate UIs, complete forms, interact with APIs, analyze data, and execute workflows.
  • Difference from classical RPA: Agents use perception (image, context) to decide actions rather than brittle rules.

6. Deterministic vs. Probabilistic Components

• ** task allocation:**

  • Deterministic systems (e.g., for math, analytics) handle what LLMs can’t reliably do.
  • LLMs are leveraged for insight synthesis and narrative, while deterministic engines ensure reliability and audit.
    • “If you want to be absolutely sure that the math is correct…they’re doing tool calling, going into a deterministic system.” – Arijit Sengupta [19:21]

• Feedback Cycle:

  • The reasoning model enables the AI to explain its steps, allowing high-resolution human feedback (“show your work”).
  • This leads to improved outcomes with less training data and immediate user payoff.
    • “In reasoning model, when you give feedback, we immediately make the AI update itself with that feedback, redo that work.” – Arijit Sengupta [27:47]

7. Evolving AI: Human-AI Synergy & Model Specialization

• Agent Evolution:

  • New models are called “interns”—they start unspecialized, but over time develop niche expertise based on accumulated, contextual feedback from diverse users.
    • “What happens is these interns evolve into many, many specialized agents…through this kind of human feedback and evolution.” – Arijit Sengupta [27:46]

• Analogy:

  • Treat AI agents like human interns—train, monitor, and gradually give more autonomy as they gain experience, but ensure oversight and context-specific guidance.
    • “Think about as an intern… they have some training just like the models have some training and… memory, long term memory, if they have.” – Tony Ambrosi [25:02]

8. Reasoning, Creativity, and Guardrails

• Reasoning Models:

  • Forcing AI systems to articulate their logic improves accuracy and fosters learning (akin to “showing your work” in math education).
  • “If we force the models to explain what they do and how they think through things, they get to better outcomes.” – Tony Ambrosi [21:52]

• Balancing Creativity and Control:

  • Allowing agents some randomness (stochasticity) is critical for adaptability but requires careful design to avoid unintended (costly) consequences.
    • “You want to control for bad things, but also you want that agent to have the creativity to try different things.” – Tony Ambrosi [29:19]

Notable Quotes & Memorable Moments

• On Rapid Prototyping & Success:
  • “Something real has to be in the hands of business at scale in 30 days.” – Arijit Sengupta [04:51]
• On Empowering Humans:
  • “Think, how can I give every human a superpower? And then you can do crazy things.” – Arijit Sengupta [16:51]
• On Reasoning and Feedback:
  • “You could train an AI like a pet, good job, bad job, but that’s not very high resolution feedback. The AI couldn’t explain itself to the human and the human couldn’t explain the feedback to the AI.” – Arijit Sengupta [21:34]
• On Agents as Interns:
  • “Treat AI agents like an intern: they have some training… memory, long term memory, if they have… you still don’t do that [let them go unsupervised].” – Tony Ambrosi [25:02]
• On the Near Future:
  • “I don’t see another winter in AI… Given the investments… it’s now back to us to say how do we use these things in the most logical way.” – Tony Ambrosi [33:15]
• On Hyper-specialized Agents:
  • “Not one big model to rule it all, but each user benefiting from hundreds and thousands of autonomous agents assisting them almost unseen.” – Arijit Sengupta [36:39]

Timestamps for Key Segments

• [02:03] Tony’s background and meeting Arijit at Disney—beginning of their AI partnership
• [04:51] The 48-hour prototype process and why speed matters
• [06:38] How Aible’s platform and NVIDIA’s tools empower non-technical teams
• [08:21] Responsible AI practices within CVS Health and focus on customer experience
• [10:43] Defining “AI agents” and the fusion of LLM intelligence with process automation
• [13:29] The agent autonomy scale (0–5) and practical deployment realities
• [17:19] On empowering humans, not just automating tasks
• [19:21] Distinguishing between deterministic and probabilistic operations in AI systems
• [21:34] Reasoning models and the value of stepwise human feedback
• [25:02] Treating AI agents as human-like interns—training and oversight
• [27:46] The evolution from “image intern” to highly-specialized agent
• [29:19] The need to balance creativity and safety in agentic AI
• [33:15] Predictions for the next wave of enterprise AI
• [36:39] The rise of thousand-agent ecosystems and emergent intelligence

Final Thoughts & Future Implications

Both guests agree that AI’s current pace of progress is exponential rather than linear, driven by feedback loops and agent specialization. Enterprises will not rely on one massive model but on ecosystems of thousands of specialized, continuously evolving agents. The goal is not just automation, but human empowerment—augmenting users and business decision-makers so they can achieve previously impossible goals.

• “The more intelligent agents you have in the mix, the better the entire result is… sometimes the agents themselves negotiate with each other and each other's reasoning and that's an emergent capability.” – Tony Ambrosi [36:56]

Businesses should thus focus on:

• Rapid prototyping and iteration
• Maintaining a “human in the loop” approach
• Investing in explainability, feedback, and reasoning
• Anticipating a world of specialized, adaptive, agent-driven AI systems

Additional Resources

• Aible: aible.com — includes blogs and partnership stories with NVIDIA [38:12]
• NVIDIA AI Podcast: ai-podcast.nvidia.com

This summary covers the key discussions, insights, and memorable moments from NVIDIA AI Podcast Episode 261, designed to provide clear value for those who haven’t listened to the episode.