Aakash GuptaAI Agents for PMs in 69 Minutes — Masterclass with IBM VP
CHAPTERS
- 0:00 – 0:14
Defining AI agents: from chatbots to the “wall of automation”
Aakash asks what makes an AI agent different from ChatGPT-style chatbots. Armand frames agents as the practical realization of AI’s promise: automation that meaningfully increases business output, especially in enterprise settings.
- •Agents as the next step beyond predictive analytics and chatbots
- •Why agents unlock automation and higher organizational throughput
- •Early agent examples (AutoGPT-style) vs today’s renewed focus
- •Enterprise demand: agents as a top CIO priority
- •Production realities: experimentation vs safe, secure deployment
- 0:14 – 7:06
The four-step agent loop: Think → Plan → Act → Reflect
Armand explains his core mental model for agents using four phases. He emphasizes that action and reflection are what turn an LLM from a helpful assistant into a system that executes and improves over time.
- •Thinking/reasoning requires more inference tokens but improves outcomes
- •Planning decomposes goals into subtasks and can challenge prior outputs
- •Action connects the model to real systems (CRM, email, HR tools)
- •Protocols like MCP expand what actions agents can safely perform
- •Reflection enables iterative improvement via history + human input
- 7:06 – 9:35
How to build agents: coding frameworks vs low-code/no-code builders
They map popular agent development options into two buckets: programmable frameworks and visual builders. The discussion focuses on choosing the right level of control depending on complexity and production needs.
- •Coding frameworks: LangGraph, CrewAI, LlamaIndex, AutoGen
- •Low/no-code: LangFlow, Lindy, n8n, Stack AI, Flowise (and similar tools)
- •When you need “full control” vs when speed and accessibility matter
- •Open-source velocity: learning from GitHub issues/PRs and community iteration
- •The importance of understanding concepts even when using no-code tools
- 9:35 – 11:09
RAG explained: adding fresh enterprise context to LLMs
Armand clarifies what Retrieval-Augmented Generation (RAG) is and when it’s preferable to fine-tuning. He highlights how central RAG has been for enterprise AI, especially right after ChatGPT’s release.
- •RAG injects up-to-date context into models trained on static snapshots
- •Fine-tuning vs RAG: different goals (behavior/skills vs changing knowledge)
- •Why RAG dominated early enterprise implementations (the “90%” claim)
- •RAG as a goldmine for companies with large stores of internal data
- •RAG as a foundational component inside agent planning workflows
- 11:09 – 18:26
Enterprise RAG in practice: from search to actionable answers (incl. Vision RAG)
They explore how RAG shifts enterprise search from metadata and links to direct, usable information. Armand introduces Vision RAG for extracting meaning from PDFs, tables, and charts—unlocking multimodal enterprise data.
- •RAG as a pipeline step embedded in agentic planning
- •Use cases: pulling insights from documents + structured sources for decisions
- •Why this changes PM work: rapid access to customer/usage/feedback intelligence
- •Key RAG building blocks: embeddings, vector DBs, filtering/ranking
- •Vision RAG: multimodal extraction from PDFs/tables/charts (e.g., IBM DocLink)
- 18:26 – 22:27
Common RAG failures: accuracy, “vanilla” implementations, and why evals must be everywhere
Armand outlines why teams struggle with RAG in production—primarily accuracy and trustworthiness. He argues evaluation shouldn’t be limited to the final response; it must be applied throughout multi-step agent workflows.
- •Enterprise tolerance for errors is far lower than consumer apps
- •Why off-the-shelf RAG templates often fail real production requirements
- •Evals should be inserted at multiple steps, not just at the end
- •Defining acceptable business accuracy and iterating architecture/configuration
- •Treating many AI failures as data engineering and data quality problems
- 22:27 – 25:23
Scaling evaluation: empowering SMEs and building organizational governance for AI cost/impact
The conversation moves from technical evals to operationalizing them across large organizations. Armand emphasizes SME involvement, continuous monitoring, and balancing innovation against compute cost and ROI.
- •Methods: synthetic data, ground-truth datasets, and human review loops
- •IBM’s Eval Studio concept: GUI-based evals for business users/SMEs
- •Continuous evaluation and maintenance as an ongoing operating model
- •Enterprise-scale frameworks: AI hubs/AI offices to track initiatives
- •Avoiding “$20k compute to get $100 benefit” productivity illusions
- 25:23 – 28:52
Managing 10–20 agents per employee: orchestration, bottlenecks, and human accountability
Armand describes a future where individuals “manage” many specialized agents like a team. They discuss orchestration as the emerging skill: coordinating agents, setting approvals, and owning outcomes.
- •Start from functional bottlenecks (e.g., marketing copy, creative, A/B testing)
- •Specialized agents for recommendations vs full execution automation
- •Orchestration as a core job: iterate quickly, validate outputs, ensure safety
- •Human-in-the-loop approvals to protect brand and reduce operational risk
- •Adoption gap: AI-native startups vs slower-moving traditional enterprises
- 28:52 – 35:15
How AI reshapes product management: higher PM leverage across the lifecycle
They focus on what PMs should build first and how PM scope changes with agents. Armand predicts PM-to-engineer ratios shift as agents take over research, synthesis, drafting, and monitoring tasks.
- •Potential shift from ~1 PM per 6–10 devs to 1 PM per 20–30 devs
- •PM agent ideas: competitive intelligence, market monitoring, sales enablement
- •Synthesizing feedback: usage data + NPS + social/qual signals for prioritization
- •PRDs drafted 80–90% by AI, with PM refinement
- •Post-launch monitoring agents detect edge cases and summarize outcomes
- 35:15 – 39:04
Prototype-first vs “feature factory”: keeping customer discovery central
Aakash challenges the risk of rushing from idea to prototype without real problem investigation. Armand argues prototyping accelerates communication, but customer-first discovery remains essential to avoid shallow solutionism.
- •Armand’s career story: prototypes beat slides/PRDs in exec decision-making
- •Why prototypes reduce lost-in-translation issues (global teams, language barriers)
- •Risk: AI prototyping can push teams into a feature factory mindset
- •Correct sequence: deep customer discovery, then rapid prototyping to validate
- •PM craft still requires problem understanding, not just fast building
- 39:04 – 44:52
Upskilling roadmap: technical literacy, build one agent, then deepen with vibe coding + Python
Armand advocates technical literacy for all roles to grasp the “art of the possible.” He offers a practical progression: learn concepts, build a simple agent with accessible tools, then level up into coding and deeper system understanding.
- •AI as growth/enabling tool, not only cost savings (referencing Aaron Levie)
- •Hands-on experimentation beats passive learning (posts, videos, papers)
- •Step 1: learn core concepts (LLMs, reasoning, RAG)
- •Step 2: build one agent with a no-code/low-code flow builder
- •Step 3: try vibe coding, then Python for advanced experimentation
- 44:52 – 48:13
Can open source win? Why enterprises prioritize control, deployment flexibility, and ecosystems
They debate open vs closed models through an enterprise lens. Armand argues open source ‘always wins’ long-term for enterprises due to deployability, governance needs, and the broader open ecosystem (frameworks, infra, tooling).
- •Enterprise constraints: data confidentiality and tool integration
- •Deploy anywhere: on-prem/private cloud/hybrid without reliance on API calls
- •Community pace: open source catches up over time even if behind initially
- •Ecosystem advantage: Kubernetes, vLLM, PyTorch, and OSS agent frameworks
- •Why licensing and governance matter as much as model quality
- 48:13 – 51:39
Open-source milestones and model-building basics: Hugging Face, Mistral, Llama, PyTorch, and training phases
Armand reviews notable open-source releases and how innovation spans every layer of the AI stack. They briefly demystify pre-training, post-training, and alignment, plus the increasing role of high-quality synthetic data.
- •Open-source LLM highlights: Mistral (MoE), Llama ecosystem, IBM models
- •Hugging Face as a hub: hundreds of thousands of models + datasets
- •PyTorch’s dominance (vs TensorFlow’s earlier promise) and why it matters
- •Training pipeline basics: pre-training, post-training, alignment
- •Data scarcity and the rise of synthetic data supervised by humans
- 51:39 – 56:31
IBM’s AI strategy: deploy-anywhere platform, Granite models, and governance for regulated scale
Aakash presses on how IBM will ‘make waves’ in AI, and Armand lays out IBM’s enterprise-focused strategy. He emphasizes hybrid deployment flexibility, cost control, model governance, and small efficient models (Granite) with transparent data provenance.
- •Platform strategy: flexibility across hyperscalers, on-prem, private cloud, hybrid
- •Customers want multiple model options plus governed access and cost visibility
- •Scaling inference across clusters and environments
- •Granite model value: small, cost-efficient, easier to customize
- •Governance + compliance tooling: inventory use cases, meet evolving regulation, manage risk
- 56:31 – 1:09:19
Career + influence: intern-to-VP lessons and building 200K followers with (selective) AI assistance
Armand shares how intentional moves, timing, and consistent execution shaped his rise from intern to VP. He also explains his LinkedIn growth system, including how he used AI more heavily early on but now relies more on original thinking and targeted audience building.
- •Intentional path to Silicon Valley and consistent focus on AI through ‘winters’
- •Career accelerants: customer closeness, networking, shipping fast, ‘show not tell’
- •Why AI research talent commands extreme compensation (capital allocation context)
- •Content system: daily posting routine, idea capture, writing as thinking
- •Using AI for research/ideation early; reducing AI use later to avoid homogenized content
