Lore

AI product management

How Freshworks' CPO Actually Builds Products With AI (Live Demo)

Freshworks CPO Shini Ragaban argues that AI has killed the traditional linear PM→UX designer→engineer handoff, merging those roles into a single "product builder," and he demonstrates this live via Freshworks' internal "AI PDLC" pipeline (Cursor, PRD Genie, Figma Make, Agent Studio, and a Claude-MCP connector) that he says took the company from a six-month release cycle to a two-week one.

Aakash Gupta · 2026-08-24 · English

Key ideas

  1. Problem framing: PM/UX/engineer titles existed only because of a linear handoff (PRD → Figma prototype/wireframe → engineering build); AI removes the need for that handoff, so the three roles are collapsing into one "product builder" role, already visible at some startups and predicted to spread widely.

  2. Foundation-first claim: an "AI-first" approach without groundwork is a mistake; establishing a data/design foundation first (design system, coding system, reusable component repository) is what makes subsequent AI-first work faster, and this belief underlies Freshworks' whole AI PDLC architecture.

  3. Freshworks operationalized this belief as a governed, company-specific AI PDLC spanning discovery through release, with an AI agent assisting at every phase, justified by the claim that at their scale (75,000 customers, ~300 million end users) an ungoverned or hallucinating AI is unacceptable.

  4. Demonstration as argument: instead of describing the system abstractly, the speaker performs a full live walkthrough (problem statement → PRD Genie-drafted PRD → Figma Make prototype → live Agent Studio demo → MCP-driven ticket automation) to substantiate the claim that AI collapses both roles and cycle times.

  5. Human judgment is repositioned rather than eliminated: AI functions as a "co-pilot," and the value of the human shifts from doing operational drafting work to exercising judgment — choosing correct data references, catching design/UX gaps the AI misses, and preventing hallucination through explicit grounding.

  6. Organizational consequence claimed: this shift has already changed team composition (PM-to-engineer ratio moving from roughly 1:20 to near 1:1, with some teams lacking a dedicated UX designer) and enabled release cadence to compress from six months to two weeks, with a further prediction of shrinking to every few days within six months.

  7. Counter-argument acknowledged and answered: shipping faster doesn't automatically improve outcomes, since historically about 90% of shipped features don't move metrics; the claimed real gain is faster learning and failing fast, to be judged against a "three-thirds" view of the product job (launch, adoption, monetization) rather than shipping speed alone.

  8. Change-management claim: transforming a product org this way requires leaders to personally build and demonstrate the new workflow (leading by example via recurring internal showcases) rather than mandate it top-down, and hiring/interviewing must correspondingly shift toward assessing demonstrated building ability and curiosity over credentials.

  9. Forward-looking claim: the MCP-based Claude/Freshservice integration illustrates a category shift toward software built for AI agents to consume, not just humans, meaning product builders must "unlearn" human-only design assumptions and design for both audiences going forward.

  10. AI PDLC (AI Product Development Life Cycle) — Freshworks' end-to-end AI-assisted process covering discovery, design, planning, development, QA, deployment, and release, with an AI agent assisting at every step and governed by an eval framework. Apply: Structure product development as a discrete lifecycle where each phase has AI agent assistance layered on top, grounded by governed knowledge sources rather than ad hoc AI use.

  11. Prism — Freshworks' internal system combining a Knowledge Hub (product knowledge/dependencies), a Context Hub (feature context passed across phases), and a central skills repository of "AI builder artifacts" (skills, rules, commands, agents). Apply: Centralize product knowledge, cross-phase context, and reusable AI skills/rules/commands/agents in one governed repository so every AI step in the pipeline draws from the same grounded source.

  12. PRD Genie — An AI tool used inside Cursor that drafts roughly 80% of a PRD almost instantly, grounded in usage data, competitive benchmarks, and customer feedback. Apply: Feed a problem statement into the tool so it drafts specs and pulls supporting evidence (data-lake metrics, competitor benchmarks, customer feedback) into a near-complete PRD for a PM to refine.

  13. CPO check — An AI agent step that reviews PRD documents for strategic alignment, clarity, and edge cases, automating a review the CPO previously did personally. Apply: Run finished PRD/requirement docs through this automated review step before they proceed to design/build, as a final quality gate.

  14. 12-step / 12-phase AI PDLC process — A defined sequence — idea briefing, requirements, knowledge gathering, competitor analysis, voice of customer, metrics/data-gap analysis ("KSA hypothesis"), PRD generation, visual prototyping, QA, and a final CPO-review checklist — that produces a markdown artifact at each stage. Apply: Initialize a new feature by walking an AI agent through each named phase in order, producing a markdown artifact per phase, so nothing is skipped and every claim traces to a cited source.

  15. Data-first vs. AI-first approach — A distinction between building AI-first without foundations versus first establishing a design system, coding system, and reusable component repository before layering AI-driven building on top. Apply: Build the foundational design/coding system and reusable repository first, then apply AI-first workflows on top of it, since the foundation makes the AI-first process faster.

  16. Cursor (IDE) — A no-code-friendly IDE used as the hub where PMs, designers, and engineers all work, chosen for being usable by non-technical people, offering flexible model choice, and supporting third-party integrations like the Figma MCP plug-in. Apply: Use Cursor as the single workspace where a "product builder" drafts PRDs, inspects code, and connects to design tools via chat/CLI rather than routing work through separate specialist tools.

  17. Figma Make + Figma MCP plug-in — A prototyping tool preloaded with the company's own design system ("Due") and product scaffolding (e.g., the Fresh Service starter) that reads Figma mockups and PRDs to generate working visual prototypes. Apply: Initialize Figma Make with the existing production scaffolding and design system, paste in the PRD, and iteratively prompt it (specifying components, responsiveness, chart types) rather than expecting a one-shot result.

  18. Sankey chart (as a prompted visualization) — A chart type explicitly requested for the performance-dashboard prototype to show proportions of ticket topics/categories (an audience comment notes what appeared was actually a treemap). Apply: Specify the desired chart type directly in the prompt to a design-generation agent when a dashboard needs to show categorical proportions, and verify the rendered output actually matches the requested type.

  19. Agent Studio "Knowledge" ingestion — A mechanism for feeding an AI agent governed content via URL ingestion, policy/SOP documents, and categorized solution articles, plus connectors to Google, SharePoint, and Confluence, to prevent hallucination. Apply: Before launching an agent, load it with URL-ingested docs, SOPs, and categorized solution articles, and connect third-party knowledge sources so answers stay grounded rather than invented.

  20. Service catalog — A governed list of requestable services/items (e.g., Slack access, employment verification letters) maintained by IT/HR so an agent cannot fulfill arbitrary requests. Apply: Define an explicit catalog of what an agent is allowed to fulfill, and restrict its actions to that catalog rather than open-ended requests.

  21. Workflows (Agent Studio building blocks) — Reusable automation building blocks (e.g., creating a PagerDuty incident, resetting an Okta/Azure AD password) composed into an agent, with common ones pre-built from analyzing the most-used 80% of customer workflows. Apply: Start from pre-built workflows covering the most common cases, then build custom workflows (e.g., a Jamf-based "lock computer" flow) for company-specific needs.

  22. MCP (Model Context Protocol) connector — A governed connector (demonstrated between Freshservice and Claude) letting an external AI assistant query product data directly, e.g., fetching tickets and generating a report, without the data leaving the environment. Apply: Connect an external AI assistant to the product via an MCP connector so a single prompt can fetch data, generate a report, and act on it (e.g., replying to tickets) instead of requiring multiple manual steps.

  23. "Hire to retire" agent — A named AI agent template covering the full employee lifecycle (onboarding through offboarding), deployed via Slack/Teams to handle IT and HR requests. Apply: Design an agent scoped to a full lifecycle journey rather than isolated tasks, and deploy it in the channels employees already use instead of a separate portal.

  24. Dual-channel deployment pattern — Making the same agent accessible via both a private channel (e.g., "Concierge") and a public channel (e.g., "HR help"), with the choice left as customer preference. Apply: Offer both private and public access points to the same underlying agent/knowledge so employees can choose their preferred level of visibility for sensitive questions.

  25. "Co-pilot, not autopilot" framing — The claim that AI performs the operational drafting work but a human still supplies judgment (choosing references, catching gaps, ensuring grounding) rather than the process running fully autonomously. Apply: Keep a human reviewer in the loop to select which data/version the AI should be grounded in and to catch what the AI misses, even as most drafting work is automated.

  26. "Product builder" role — A proposed merged role replacing the traditional PM/UX designer/engineer split, on the claim that AI now lets one person research, design, and build/ship. Apply: Reorganize hiring and team structure around a single "product builder" who owns research through shipping, rather than routing work through separate PM→designer→engineer handoffs.

  27. Three-thirds framework of the product job — Dividing the product job into launching the product, driving user adoption, and monetization, used to argue that shipping faster with AI doesn't remove the adoption/monetization work. Apply: Evaluate AI-accelerated shipping against all three thirds (launch, adoption, monetization) rather than treating faster shipping alone as success.

  28. Git-repo / live-demo interview technique — Replacing purely verbal "tell me about a time" interview questions with asking candidates to open Cursor and show what they've built, plus reviewing their git repository. Apply: For hiring in this new model, ask candidates to demonstrate live in a tool and review their own repositories rather than relying solely on behavioral Q&A.

  29. Monthly product team showcase meeting — A recurring internal demo session where team members show things they're building, cited as the mechanism by which grassroots tools (PRD Genie, the Cursor-based process) were discovered and standardized org-wide. Apply: Run a regular cross-team demo forum so individually built tools/workflows can be surfaced and promoted to org-wide standards, rather than mandating adoption top-down.

  30. SLA vs. ELA (Service Level Agreement vs. Experience Level Agreement) — A reframing of what gets measured, shifting from ticket-resolution SLAs to a broader "experience level agreement" tracking employee satisfaction. Apply: Track experience-level metrics (employee/customer satisfaction) alongside or instead of traditional SLA metrics when AI agents change how service is delivered.

  31. "Show it, don't say it" leadership approach — The claim that transforming a team's practices requires leaders to personally build and demonstrate the new way of working themselves, since staff dismiss slide decks/mandates as "ivory tower preaching.". Apply: As a leader wanting the org to adopt a new workflow, build something visible in it yourself and share it, rather than mandating change via policy documents.

Insights

The live demo itself surfaced a nontrivial catch: the AI-generated PRD flagged that employee experience (EX) uses a different "service catalog" concept than customer experience (CX), showing that grounding in company-specific data can catch domain nuances a naive reuse of an existing feature would miss.

Prototyping in Figma Make was explicitly not one-shot: the tool initially misused design-system components, broke on narrow monitors, and failed to render the Sankey chart correctly — the "judgment" value-add the speaker describes concretely consisted of catching these specific gaps through iterative prompting, giving a tangible referent for "co-pilot, not autopilot."

The governance mechanism for trust at enterprise scale isn't described as better AI capability but as explicit provenance: every generated markdown artifact cites which reference/data version it drew from, directly answering the host's raised concern about stale or ambiguously-defined metrics being silently baked into an AI-drafted PRD.

The organizational effect claimed (PM:engineer ratio compressing from ~1:20 to ~1:1, sometimes with zero dedicated UX designer) is presented as a concrete before/after data point tied to an 18-month timeline, not a hypothetical projection, directly linking the tooling narrative to a headcount/structure claim.

Interview practice at Freshworks was changed to match the skill now being valued: candidates are asked to open Cursor live and show their git repository rather than answer behavioral questions, operationalizing the "judgment and curiosity over credentials" argument at the hiring level rather than leaving it abstract.

An audience comment flags a labeling slip in the source itself (the chart shown was a treemap, not a Sankey diagram, despite being called one throughout) and separately argues the video only shows the surface: the harder groundwork of building the design system, feeding tech-stack context, and initializing project context is not revealed.

«I think these three different roles are going to be replaced by what I call as a product builder.»

— 00:04

«This is a very linear process and this is completely dead.»

— 00:36

«It all sort of starts and ends in cursor.»

— 06:34

«that's exactly the problem that the PRD Genie has solved. It drafts 80% of the PRD almost instantly.»

— 08:02

«Imagine it would take days and days to do this. The agent just... did this in a matter of... seconds, right? Not even minutes.»

— 22:19

«it's actually a co-pilot, there is still value for the product manager if the product manager is doing it for the product manager to do it»

— 26:08

«you can't afford AI to hallucinate in building software. It's just not acceptable because then 75,000 customers are going to get affected.»

— 26:41

«These AI agents cannot hallucinate, which means they need to have guardrails and they need to be governed.»

— 40:01

«What we just saw with MCP was that was not a system that was built for humans. It was built for an AI agent... to be more effective.»

— 56:29

«I would be surprised if in the next 6 months we go down from 2 weeks to maybe releasing things every 2 days every week.»

— 58:36

«I almost think of the product's job as a third of it is launching the product. The other third is actually making sure the user adopt it. The last third is monetizing it.»

— 61:47

«why don't you open a cursor and just show me what you did.»

— 63:32

Reception

Comments are largely positive and appreciative, with constructive feedback and enthusiasm but tempered acknowledgment that the video only scratches the surface.

This is a persuasive, internally-sourced case study — every example, metric, and tool is Freshworks' own — presented as evidence for a broader industry claim about role convergence; it is valuable as a concrete workflow walkthrough but should be read as one company's self-reported numbers and demo rather than independently verified industry data.

65:45

↳ Aakash Gupta · YouTube

Watch original