Tanu Mutreja

I like to think about AI roadmaps and investments through two lenses: How AI powers our business i.e. using AI for efficiency, automation, and enhancing existing products. How our business powers AI i.e. leveraging our unique domain expertise to build differentiated, defensible AI capabilities. Beyond that, while core PM principles remain unchanged, AI has introduced several new critical considerations for roadmaps and investments: Natural language and conversational interfaces are now table sta ...Read More

Technically, the biggest hurdle in shipping agentic AI today is quality. Most agents work well for narrow or simple tasks, but they struggle to reliably solve complex, end-to-end use cases. This means teams need to set realistic expectations and make deliberate design and tooling choices that complement current agent limitations rather than fight them. The second major challenge is evaluation, building systems that continuously verify an agent is behaving as intended as prompts, tools, data, and ...Read More

Yes—100%. At this point, almost all roles clearly see AI as a major productivity multiplier, not just a feature layer or an experimental bet. The conversation has shifted from “How can AI help individuals?” to “How does AI change the way product work flows end to end?”   Across product, design, engineering, and go-to-market, AI is already raising the baseline of execution. Tools like Microsoft Copilot, Atlassian Rovo, Zoom Companion and Gong are good reference points, but the bigger shift is cul ...Read More

In an AI-powered marketing landscape, the best product marketers will combine prompting, rapid prototyping, real-time feedback and sentiment analysis, cost awareness, and modern content creation skills. PMMs will increasingly become hands-on builders and artists, not just narrators. They’ll move faster by turning ideas into working experiences, using real-time analytics and user signals to iterate messaging and positioning quickly and credibly. Prompt engineering has already become a foundationa ...Read More

For testing and validating the AI components of our product, we use a multi-layered strategy that combines rigorous pre-deployment evaluation with continuous post-deployment monitoring, because AI systems are non-deterministic and evolve with data and usage.   We validate models through offline evaluation using curated test sets, golden datasets, and regression benchmarks to validate accuracy, robustness, and failure modes before anything reaches users. Note that we do not rely on public benchma ...Read More

As AI agents become more capable, the real question is where autonomy should stop and human judgment should begin. In practice, the guiding principle is straightforward: the more critical an action, the more human-in-the-loop or user control is required. Today, AI agents can perform junior-level, low-risk tasks autonomously. But as agentic actions become more critical, complex, destructive, or irreversible (e.g. code changes, system updates, record deletions, or security actions) progressively s ...Read More

A few companies are doing agentic AI particularly well. They are building agents that may not be the most autonomous, but are genuinely helpful, observable, constrained, and reviewable. My personal favorite is OpenAI agents. I use these agents every day to solve real problems, for example, turning information from web pages into actionable workflows, tracking specific news, generating summaries, and automating follow-ups. What stands out is not just model quality, but agent reliability, ease of ...Read More