In our latest episode of Deployed, we sat down with Ajit Varma, VP of Product at Mozilla Firefox, to discuss how Mozilla is navigating the integration of AI into the browser while maintaining their commitment to user privacy and an open web. It’s a thoughtful conversation that makes you glad people like Ajit are leading the charge with major tech products.

Ajit brings a remarkable breadth of product leadership experience to this conversation. Over his 20+ year career, he's been a product leader at Google (search & ads quality, Gmail, Google Calendar), founded his own AI ecommerce startup, built new products at Square, and most recently spent six years leading monetization for WhatsApp at Meta. At each stop, he's focused on identifying inflection points where technology enables fundamentally new user behaviors.

Now as head of product for Firefox, Ajit sees another such inflection point: how people use browsers is about to change dramatically with AI. But unlike Chrome (tied to Gemini), Edge (tied to Copilot), or the new Atlas browser (tied to ChatGPT), Firefox is taking a different approach: one that prioritizes user choice, privacy, and a strong commitment to keep the web open.

In this conversation, we explore:

• Mozilla's unique role and opportunity in preserving an open web as AI reshapes browsers

• How the Firefox team is building AI features that honor user privacy through on-device models and private hosting

• The tactics they use to ensure quality when they refuse to look at user data

• Practical lessons from dogfooding AI features at scale

The episode is live on Spotify, Apple Podcasts, YouTube, and you can watch the whole thing below. Read on for some of our favorite highlights.

The Future of the Open Web in the Age of AI

One of the most compelling parts of our conversation was Ajit's perspective on why Mozilla's independence matters more than ever as AI becomes central to browsing.

"If you look at over the last say, twenty years for Firefox, many of the biggest companies today that are leading AI forward are made possible because they had a distribution channel that didn't have gatekeepers. Google was a website. Facebook was a website. Even ChatGPT was a website, and people went to that website to start."

But Ajit points out that open access is a question mark when it comes to AI in the browser. Every major browser except Firefox is tightly coupled to a specific AI provider: Chrome and Gemini, Edge and Copilot, and Atlas is literally called "ChatGPT Atlas."

"If every browser only allows you to access their own AI, then you're only going to have a handful of AI's that exist... You're gonna want to control more than just a chat. You're gonna want to control almost every layer of agents."

Firefox's approach? Give users choice, and lean into their strengths. Among those strengths, Ajit talks about how the uniqueness of the Gecko rendering engine (the open source core of Firefox) makes it fundamentally different from so many other Chromium-based browsers, and how that lets them think differently about future user experiences.

Ajit also highlighted how Firefox's extension framework is uniquely powerful for AI builders. Unlike Chrome's Manifest V3, which restricts what extensions can do (ostensibly for security, but also limiting ad blockers), Firefox maintains both security protections and the full extensibility of previous frameworks. "You can build much more powerful AI extensions in Firefox than any other browser," Ajit shared.

Privacy-First AI: Making Hard Tradeoffs

Lots of folks building in AI will likely appreciate the conversation around how Mozilla thinks about privacy in their AI features. They face unique constraints that most companies don't: they simply won't log all user queries and use that data to improve their models.

This creates interesting technical and product challenges. Mozilla's approach involves multiple strategies:

• On-Device Models When Possible: For their "Shake to Summarize" feature on iOS, Mozilla uses Apple Intelligence models when available (iPhone 15 and above with iOS 18). This keeps all data on-device and maximally private.

• Private Cloud Hosting: When on-device models aren't available (older devices, Android, desktop), they fall back to hosted cloud models. But crucially, these are open-source models hosted in Mozilla's private cloud with strict data protections. No data is used for training or shared with third parties.

• User Choice as a Core Principle Looking ahead, Mozilla envisions giving users explicit choices: "Do you want to use on-device? It just means it's gonna be less capable. Or do you use cloud if you want this? And then even model choice: do you want a private open model or a closed model?"

Ensuring Quality Without User Data: Dogfood Heavily

This privacy-first approach creates interesting challenges for quality assurance. Most AI companies improve their products by analyzing production logs at scale. Mozilla can't do that.

Given their constraints around user data, Mozilla has leaned especially hard on internal dogfooding to discover quality issues. Ajit shared a perfect example of why this matters:

"Dogfood is usually the best first way to understand how people are using it. So for 'Shake to Summarize' as an example: An engineer was summarizing a recipe site, and they found that it dropped ingredients. Then you have to say, okay, do you need a longer context window? You have to look at more characters on the page."

It's a great reminder that even the most sophisticated evaluation systems need real-world usage to uncover edge cases. The recipe example highlights how domain-specific quirks (recipes having ingredients at different locations, long preambles, etc.) might not be obvious until someone actually tries to use a feature.

For anyone thinking about these sorts of feedback loops: These kinds of discoveries should then be used to update your evaluation datasets. The recipe issue for instance can become a test case you check against every time you change models or update prompts. Over time, this sort of dataset curation builds a comprehensive evaluation suite that captures real failure states, not just imagined ones.

Dogfooding isn't the only path when privacy is a concern. Another tip from a prior guest: Ben Kus, CTO at Box, described how they use LLM as a judge to evaluate quality without human review of sensitive customer data. The LLM judge looks at log data and produces a score that the Box team can use without looking at the data itself. Check it out here.

Building with AI for the Right Reasons

We wrapped our conversation by asking Ajit for advice to other teams building AI products. His response cut straight to values and purpose:

"I'm one of the AI optimists. And so I think there's a lot of unknowns, but if you're optimistic about the future, I think what you should build is things that assume AI is gonna be good, is gonna be transparent, is gonna reduce cost. And then just make sure that you embody those values in the products that you're creating and that you're using AI for positive outcomes for humanity - making humans more productive and happier and breaking down barriers."

It was refreshingly human advice from a seasoned product leader, and an emphasis we could probably all use more of. Just because we can doesn't mean we should. As AI becomes more powerful and pervasive, we need more teams thinking seriously about the values they're embedding in their products and the futures they're creating.

Conclusion

For teams building AI products, Mozilla's journey offers valuable lessons: be thoughtful about the values you embed in your product, the ways you approach user privacy, and look for ways to differentiate through your user experience in ways that competitors can't.

Thanks to Ajit and the Mozilla team for sharing these insights. For more on Firefox's AI features, visit firefox.com.

Want to hear more conversations like this one? Subscribe to Deployed on Spotify, Apple Podcasts, or YouTube.