Aakash GuptaWhat AI PMs REALLY Need to KNOW in 2026 (Agents, Discovery, EVERYTHING)
CHAPTERS
AI PM job boom and the danger of “AI” as a label
Aakash opens with data showing AI PM job postings doubling year-over-year, and Todd explains why job posts are partly marketing signals. Todd warns that calling yourself an AI PM invites extra scrutiny because of rampant “AI washing.”
- •AI PM job postings rising from ~10% to ~20% of PM listings
- •Job postings reflect both hiring need and employer branding
- •AI is hot across industries, so “AI PM” demand is unsurprising
- •Caution: claiming “AI PM” requires real competence and proof
Upskilling mindset: using AI at work vs building AI into products
Todd separates two AI tracks PMs must master: personal productivity and shipping user-facing AI features. He argues that PMs who don’t use modern AI tools for research/prototyping will fall behind on speed and discovery.
- •Two tracks: AI for your workflow vs AI embedded in customer product
- •Prototype faster with tools like Replit/V0/Bolt/Lovable to reduce dependency bottlenecks
- •Use AI for competitive/market research (e.g., “deep research”)
- •Even ‘non-AI’ features may use LLMs under the hood—don’t over-market it
Why AI PMs earn 30–40% more (and the resume “AI washing” trap)
Todd explains AI compensation premiums as a mix of market heat and skill scarcity, similar to other deep technical PM specializations. He stresses that employers will “interrogate” AI claims because fake credentials and superficial exposure are common.
- •Pay premium driven by scarcity + perceived leverage of AI productivity gains
- •Analogy: analytics/technical PMs often paid more than UI-oriented PMs
- •Employers heavily validate real production experience
- •Warning against shallow certificates and inflated claims
The upskilling roadmap: a 5-layer “technical pyramid” for AI PMs
Aakash lays out a step-by-step competency pyramid: fundamentals first, then observability/cost, then evals, then roadmap/stakeholders, and finally ethics/leadership. This frames the rest of the episode as a structured learning path.
- •Layer 1: AI/ML basics, data pipelines/RAG, prompt engineering
- •Layer 2: observability/trace analysis, monitoring, cost optimization
- •Layer 3: evals, QA, metrics and experimentation
- •Layer 4: product strategy/roadmap and stakeholder management
- •Layer 5: leadership, ethics/safety, culture/team building
AI/ML fundamentals: model choice, token economics, privacy, and constant change
Todd emphasizes hands-on experimentation and understanding model tradeoffs (quality, speed, cost). They discuss multimodality, open-source/self-hosting, data residency, and how vendor constraints shape product decisions.
- •Know model tradeoffs: quality vs latency vs cost; when to use different tiers
- •Token economics and context windows as practical constraints
- •Gemini’s multimodal strengths (e.g., video) as an example of model fit
- •Open source/self-hosted models for privacy, security, and procurement simplicity
Data pipelines & RAG: getting the right context at scale
They argue data pipelines belong in the foundation because most real products need RAG-style context injection. Todd explains embeddings, vector databases, and the performance/governance challenges of shipping this reliably at enterprise scale.
- •RAG as a default architecture for many AI product features
- •Embeddings + vector DB retrieval to supply relevant context to LLMs
- •Too much context can degrade answers—precision matters
- •Scale, latency, and governance/data residency are key design constraints
Prompt engineering: not hype—instruction quality and platform vs domain PM roles
Todd frames prompting as the skill of precise instruction and contextual setup, analogous to being good at search. They predict further specialization (AI platform PMs enabling domain PMs) while still requiring broad prompt fluency.
- •Better prompts = better instruction + context, improving output quality
- •Prompting is a durable skill even as tools evolve
- •Likely role split: AI platform PM (infrastructure) vs domain PM (use cases)
- •Research PM vs product PM separation already emerging in AI orgs
Trace analysis, agents, and PM–engineering boundary lines
As orchestration grows (agents calling tools/other agents), trace-level understanding helps diagnose failures and performance issues. Todd notes real tension: some engineering leaders resist PMs “shadowing” technical debugging, so PMs should be fluent without overstepping.
- •Agents increase complexity: chains of calls, tool usage, recovery loops
- •Trace analysis helps locate breakdowns inside long orchestration flows
- •PM should know enough to ask ‘what do the traces say?’
- •Ownership boundaries vary by company size and culture; partnership is key
Production monitoring realities: ops/SRE, access controls, and company context
Todd explains that monitoring is often owned by ops/SRE teams, and PM access may be restricted by contracts and background-check requirements. The takeaway: understand monitoring concepts, but adapt expectations to your org’s operational model.
- •Monitoring is often centralized (ops/SRE), not owned by feature teams
- •PM depth depends on role (platform PM vs feature PM) and company maturity
- •Customer data access often limited contractually; not everyone can “touch prod”
- •Quality/bugs still impact roadmap capacity, so PMs must appreciate tradeoffs
Cost & performance optimization: COGS, gross margins, and the path from speed to efficiency
Todd argues AI economics will increasingly matter as companies must reach sustainable margins. They discuss how early builds over-optimize for speed, then require re-architecture (caching, smaller models, tuned systems) to reduce costs and improve latency.
- •AI features directly affect COGS and long-term gross margin viability
- •Current market tolerates low margins, but sustainability will force optimization
- •Smaller/tuned models + caching strategies can preserve quality at lower cost
- •Performance wins often reduce cost because compute consumption drops
Evals are the PM’s domain: AI QA, metrics, and experimentation cadence
Todd says eval design and management is where PMs should lead—AI grading AI requires product judgment about quality and outcomes. They also cover how AI lowers the cost of variants, making experimentation more mandatory and more frequent.
- •Evals differ from classic QA: PMs best define success and test sets
- •Engineers provide harness/frameworks; PMs author and prioritize eval coverage
- •A/B testing fundamentals still apply; AI makes variants cheaper and faster
- •Shift toward outcome-based metrics (e.g., tickets resolved) over activity metrics
AI product roadmap & discovery: solve hard problems, avoid shiny objects, and kill weak features
Roadmapping starts with hard, high-leverage workflows and unique data/context advantages—otherwise you’re just “wrapping ChatGPT.” Todd stresses rapidly sunsetting low-retention AI features and building a distinct point of view (workflow-centric vs agent job titles).
- •Prioritize workflows that are tedious, high-value, and uniquely informed by your product data
- •Avoid AI-for-AI’s-sake; don’t ship a thin wrapper over generic LLM behavior
- •Be willing to turn off weak AI features quickly to protect trust and adoption
- •Strong POV matters: workflows over org charts; modality/agent-mode vs named agents
Live demo: Pendo’s AI—agent analytics, rage prompts, dashboards, and agent mode workflows
Todd demonstrates how Pendo measures and improves agent experiences: conversation analytics, topic clustering, retention by use case, and “rage prompts” with replay context. He also shows an integrated dashboard approach and agent mode that executes cross-platform analysis with guardrails and multimodal outputs.
- •Agent analytics: conversations, prompts, unique visitors, retention by use case
- •Rage prompts + replay to detect frustration and diagnose experience issues
- •Dashboards as self-serve artifacts tying AI investment to business outcomes (ROI)
- •Agent mode supports broad workflows (qual + quant), interactive charts, and clarification questions for trust
Discovery acceleration & enterprise synthesis: customer finder, MCP, and aggregating insights across systems
The demo continues with AI-assisted discovery: identifying interview targets, generating outreach guides, and automating scheduling workflows. Todd also highlights MCP as a real integration standard and shows AI summarizing feature requests from sources like Gong, support tickets, Salesforce, and CSVs into actionable themes and linked ideas.
- •Customer finder: select interview candidates with justification, then action (segment/CSV/guide)
- •Automated in-app guides with CTAs (e.g., Calendly) to streamline recruiting
- •MCP adoption rising; enables agents to use product APIs more broadly
- •Qual synthesis across enterprise tools surfaces themes and can push into delivery systems (e.g., Jira)
Stakeholder & board management for AI: control the narrative, align to outcomes, and build for regulatory change
Todd advises proactively framing the AI strategy for boards, using them as helpers rather than approval machines. They cover aligning AI bets to business objectives and designing AI systems with toggles/guardrails to adapt across regions, industries, and evolving regulation.
- •Board questions follow your narrative—show up with a coherent strategy and evidence
- •Use boards for feedback and cross-company perspective, not just ‘approval’
- •Tie AI roadmap to shareholder value via clear outcomes and business metrics
- •Ethics/privacy/regional constraints require configurable switches and fast adaptability
