# Cemre Güngor: How to ship AI features users actually use | Product in Practice Source: https://www.youtube.com/watch?v=tj1eDd0VEmE Recap page: https://rapidrecap.app/video/tj1eDd0VEmE Generated: 2026-01-13T18:45:20.014+00:00 --- ## Quick Overview The most durable pattern for shipping successful AI features involves product managers focusing rigorously on solving customer problems rather than getting overly excited about novel AI solutions, which requires learning through shipping small, iterative versions to validate real-world value over relying on initial demos. **Key Points:** - The most stable pattern for good product teams remains starting with the end in mind and avoiding getting "too excited about solutions and not problems," a pitfall exacerbated by easy AI prototyping. - Product management craft must shift from heavily metric-driven funnels (like at Instagram) to relying more on user understanding and judgment when product surfaces become complex, such as in the Figma canvas or the browser. - The key value of an AI browser like the one at the Browser Company is not friction reduction for simple queries, but the ability to bring in and utilize the user's disparate context from various apps and documents. - The single best predictor for committing to an AI feature is whether internal teams are actually using it themselves, as demonstrated when the team pivoted a highly-praised feature because internal users were not clicking on it in production. - Building AI features requires significant upskilling, including learning prompt engineering, and successful evaluation relies on setting up evals and data sets to hill climb on quality, as non-deterministic features are hard to assess beforehand. - The Browser Company found that integrating AI features requires learning by doing; they successfully shipped Gmail/Calendar integrations only after reasoning models improved tool calling capabilities, after an initial attempt failed a year prior. - Product managers should "do the boring thing" by focusing on reliable friction points that AI can solve, rather than chasing shiny, stage-demo-worthy features that lack sustained daily utility. **Context:** This interview features Cemre Güngor, who has held roles at Facebook, Instagram, Figma, and now drives product at the Browser Company (DIA), discussing how product management must evolve to ship AI features that users actually adopt. The conversation centers on recurring patterns observed across these diverse product environments—from massive consumer social apps to specialized design tools and now an AI-native browser—focusing specifically on the challenges of context management, evaluation, and iterative development in the age of LLMs. ## Detailed Analysis Cemre Güngor emphasizes that despite rapid technological change driven by AI, the fundamental product principle of focusing on the problem first, not the solution, remains paramount, even though current tools make prototyping solutions easy. He notes a crucial shift in PM craft: moving from simple, measurable funnels common in billion-user consumer apps like Instagram to complex, varied workflows (like in Figma or the browser) where success is not obvious through metrics alone, demanding greater reliance on user judgment. Güngor detailed his prior experience as the first AI PM at Figma and his current role dealing with the open-ended problem space of the browser as an AI command center. A key learning for shipping durable AI features is the necessity of learning through shipping small, iterative versions, as seen when DIA realized the core value proposition was not the command bar router, but the 'ask on page' sidebar feature after initial user testing. Furthermore, he stresses that for non-deterministic LLM features, internal dogfooding is the best validation metric; features that looked magical on stage were abandoned internally when users stopped using them post-launch. The Browser Company's evolution required company-wide upskilling, including prompt engineering, and they follow the 'bitter lesson' by waiting for model capabilities to improve (like reasoning models enabling better tool calling for integrations) rather than over-investing in fine-tuning when simple prompting yields low quality. ### Stable Product Patterns - The most stable pattern is "starting with the end in mind" and avoiding excitement over solutions before problems are defined - Evaluating prototypes requires alignment on the problem being solved - Current AI prototyping trend is powerful for agency but requires doubling down on solving customer problems. ### Evolving PM Craft Across Surfaces - Moving from Instagram's simple funnels to Figma and the browser required leaning on user understanding because success metrics are not obvious - At Instagram, funnels like posting media were clear; at Figma, usage like adding frames does not predict designer productivity. ### AI Development and Evaluation - The best predictor for committing to an AI feature is internal usage, exemplified by pivoting a feature because internal users were not clicking it despite initial positive demos - Non-deterministic LLM features require iteration and evals; shipping the smallest possible thing first is even more critical now to avoid spinning wheels on incorrect problems. ### Context as the Biggest Pain Point - The key value of an AI browser is holding all user context from various apps, alleviating the pain point where finding relevant context for an AI query feels like more trouble than it's worth. ### Learning Through Model Evolution - DIA failed to ship reliable Gmail/Calendar integrations a year ago, but succeeded recently because reasoning models improved tool calling - The company follows the 'bitter lesson,' shelving features if simple prompting doesn't reach 70% quality, preferring to wait for better model capabilities. ### DIA Product Value Revelation - DIA initially thought the command bar router (deciding between search engine or assistant) was the core value, but user feedback revealed the AI sidebar ('ask on page') and the ability to mention/contextualize multiple tabs were the true value drivers. ### Practical AI Workflows Demonstrated - Güngor demonstrated three skills: simulating a manager's perspective by attaching their written documents as context to a prompt - Synthesizing feedback from a document review meeting by combining the document and the meeting transcript to generate a succinct TLDDR and specific revisions - A daily routine where DIA uses memory/browser history to suggest the highest-pressing work items for the day.