Aakash GuptaHow Freshworks' CPO Actually Builds Products With AI (Live Demo)
CHAPTERS
- 0:00 – 3:55
Freshworks’ bet: roles converge into a single “product builder”
Srini argues the classic PM → designer → engineer handoff is breaking down as AI enables individuals to cover research, design, and build. He frames the near-future org design as fewer specialized titles and more end-to-end product builders.
- •Traditional linear handoffs are “dead” as AI collapses workflow steps
- •A single person can research users, prototype, and ship faster
- •This shift is already happening in startups and will spread to large companies
- •Implication: teams optimize for outcomes and iteration speed, not role boundaries
- 3:55 – 4:56
Data-first vs AI-first: foundations that make AI development reliable
Before demonstrating tools, Srini distinguishes ‘AI-first’ hype from a ‘data-first’ foundation that makes AI effective. He emphasizes design systems, coding standards, repositories, and governed context as prerequisites for safe acceleration at Freshworks’ scale.
- •AI speed depends on strong underlying product/engineering foundations
- •Design systems + reusable repos reduce variability and hallucinations
- •Large-scale SaaS needs governance, evals, and repeatable process
- •Build the “rails” first, then let AI amplify velocity
- 4:56 – 6:57
The AI PDLC at Freshworks: agents across discovery→design→delivery
Srini introduces Freshworks’ AI Product Development Lifecycle (AI PDLC), where AI agents support each stage from discovery through QA and release. He explains the ‘Prism’ structure: knowledge hub, context hub, and a skills repository for commands, rules, and agents.
- •AI agents assist every PDLC step, not just coding
- •Prism components: knowledge hub, context hub, skills repository
- •Governed framework and evaluation to ensure quality at enterprise scale
- •Today’s deep dive will focus on discovery + design agents
- 6:57 – 9:58
PRD Genie in Cursor: reducing PM operational burden with grounded evidence
Srini explains PRD Genie, which drafts most of a PRD quickly while gathering customer evidence, competitive benchmarks, metrics, and dependency analysis. The goal is to reclaim PM time from data gathering and documentation toward strategic judgment.
- •PRD Genie drafts ~80% of PRD and hunts for supporting evidence
- •Pulls from Freshworks data lake (Baikal) for usage metrics
- •Includes competitive analysis and internal dependency mapping
- •Adds an automated ‘CPO check’ for alignment, clarity, and edge cases
- 9:58 – 13:04
Why Cursor: accessible IDE, model choice, and Figma connectivity
Srini justifies Cursor as the hub for their workflow because it’s approachable for non-engineers, supports picking the right model for the task, and connects to tools like Figma (including MCP plugins). This enables PMs and designers to work in the same environment where artifacts become code-ready.
- •No/low-code feel makes Cursor usable by non-technical roles
- •Model flexibility: choose fast/cheap vs best-in-class as needed
- •Integrations: Figma + third-party sources, including MCP plugins
- •Single environment to go from requirements to pull request workflows
- 13:04 – 19:44
Live demo: initializing the 12-step AI PDLC to generate a PRD fast
Srini walks through a real feature: an EX Agent Studio performance analytics dashboard. He initializes the project in Cursor, feeds a problem statement, answers clarifying questions, and shows how the system scaffolds files and executes its multi-step process.
- •Feature context: performance analysis dashboard for EX agents (Slack/Teams nuances)
- •12-step process creates directories/artifacts like a developer would
- •AI asks domain-specific clarification questions like a PM
- •Fast iteration enabled by model choice (Groq for speed)
- 19:44 – 24:42
Inside the 12 steps: SQL on the data lake, VoC, competitors, and CPO review
Srini reveals what the AI produced behind the scenes: knowledge gathering, competitor research, voice of customer synthesis, and quantitative analysis via SQL queries against Baikal. The output includes structured markdown artifacts and a checklist-style CPO review gate.
- •Artifacts include idea brief, requirements, research, metrics, dependency checks
- •AI writes SQL queries to pull real usage metrics and identifies data gaps
- •Competitor scan includes Atlassian/ServiceNow patterns and gaps
- •CPO review checklist acts as a governance and quality mechanism
- 24:42 – 27:01
Trusting AI-generated PRDs: grounding, versioning, and human judgment
Aakash challenges whether AI PRDs get definitions and product reality right. Srini explains reliability comes from a deliberate initialization step that pins references/versions and keeps AI as a copilot—shifting PM value from writing to judgment and verification.
- •Grounding: AI must reference known versions/sources to avoid hallucinations
- •Initialization encodes what to reference and why
- •AI is a copilot; humans validate definitions, edge cases, and correctness
- •Quality controls are mandatory at 75k-customer scale
- 27:01 – 34:36
From PRD to prototype: using Figma Make with Freshservice scaffolding
Srini shows how they avoid ‘blank canvas’ prototyping by starting from Freshservice’s real UI scaffolding and the Due design system. He feeds the PRD into Figma Make to generate the Analyze/Performance section with multiple modules.
- •Start with Freshservice starter kit to match production shell
- •Due design system ensures component consistency
- •Figma Make builds multiple modules from PRD-driven instruction
- •Figma chosen partly for adoption—teams already live there
- 34:36 – 37:01
Where AI still needs correction: component compliance, responsiveness, charts
Srini highlights that the first design pass wasn’t perfect—he had to nudge it to use the correct design-system components, ensure it works on narrow monitors, and fix items like a Sankey chart. He frames this as the new core skill: judgment and iteration, not manual production work.
- •AI may miss design-system rules without explicit references
- •Responsiveness across screen sizes requires experienced judgment
- •Visualization details (e.g., Sankey chart rendering) can need manual prompting
- •Human role shifts to critique, constraint-setting, and refinement
- 37:01 – 43:38
Agent Studio tour: agents, workflows, and governed knowledge sources
Srini demos the EX Agent Studio product: prebuilt IT/HR agents, workflow building blocks (Okta/Azure password resets, PagerDuty incidents), and knowledge configuration. He emphasizes governance—agents shouldn’t hallucinate and must rely on curated, connected sources.
- •Agent Studio organizes IT/HR agents and reusable workflows
- •Knowledge can come from URLs, SOP docs, solution articles, and connectors
- •Enterprise connectors: Google/SharePoint/Confluence for external knowledge
- •Guardrails and governance prevent hallucinations in enterprise contexts
- 43:38 – 49:41
Slack experience demo: HR/IT concierge, forms, and letter generation
Srini shows how employees interact with agents inside Slack—asking about benefits, retrieving forms like W‑4, and generating an employment verification letter through a guided info-collection flow. He connects this to improved employee experience and faster resolution without ‘knowing someone.’
- •Private vs public channels for employee questions
- •Answers include sources/citations from knowledge articles
- •Agents can deliver documents directly (e.g., W‑4)
- •Structured workflow to generate employment verification letters end-to-end
- 49:41 – 57:06
MCP + Claude + Freshservice: ticket analytics, RCA report, and auto-replies
Srini demonstrates Freshservice’s MCP connection to Claude: retrieving Windows 11 tickets, producing a root cause analysis with a visual report, and drafting/sending responses using KB articles. He frames MCP as shifting IT agents from manual ticket handling to high-leverage oversight.
- •Single prompt: fetch tickets + generate visual RCA report
- •Governance: data stays within the environment; controlled access
- •Claude identifies clusters, timelines, likely causes, and recommended actions
- •Bulk ticket responses can be generated and posted back to Freshservice
- 57:06 – 1:05:44
Org + hiring implications: ratios collapse, release cadence accelerates
Srini describes how AI adoption changed Freshworks’ team structure and cadence: PM-to-engineer ratios compress dramatically and releases move from months to weeks (and potentially days). He outlines what PMs must learn (building for humans and agents) and how leaders should drive change by example and champion-led demos.
- •PM:engineer ratio shifts from ~1:10–20 toward ~1:1 in leading teams
- •Release cadence: six months → two weeks, trending faster
- •PMs must “unlearn/relearn” to build for both humans and AI agents (CLI/chat-first)
- •Transformation playbook: leaders demo, find champions, showcase monthly demos
- •Hiring: prioritize curiosity + proof of building (Git/Cursor demos), plus cases by level
