Browserbase is deleting hundreds of years of busy work

When Nancy at a 50-year-old dairy company spent her mornings looking up gas prices for the trucking fleet, nobody questioned it. That's just how work got done. The company had never hired an engineer in five decades of operation.

Then AI happened.

Now Nancy's free to do more interesting things, because an AI agent handles the gas price research. It opens Google Maps, searches the route, compares prices, and tells the truckers where to stop. That agent runs on Browserbase.

Paul, the founder of Browserbase, shared this story at AWS re:Invent 2025. It captures something important about where AI is actually creating value: not in flashy demos, but in the mundane tasks that eat hours out of every workday.

The web is still locked away from AI

We've all played with OpenAI's Operator. It felt, as Paul put it, "pretty mid." Exciting as a proof of concept, but not quite there yet. The gap isn't just about model quality. It's about infrastructure.

Browserbase provides the plumbing that lets AI agents actually use web browsers. Think about everything you do in a browser: filling out forms, navigating multi-step workflows, extracting data from pages that weren't built for APIs. If AI is going to do work on our behalf, it needs access to the same tools we use. The web browser is that tool.

Running browsers in the cloud sounds straightforward until you try it. Browsers were designed for desktop hardware with plenty of CPU and memory. Making them work reliably in lightweight cloud environments, securely, at scale, is genuinely hard. As Paul described it, browsers are "almost like little nuclear bombs" that can go anywhere and do anything. You need to know where they're going, make sure they succeed, and prevent them from doing things they shouldn't.

From web scraping to work completion

The old categories for this kind of tooling were web scraping and RPA (robotic process automation). Both were notoriously brittle. Write a script to extract data from a website, and it breaks the moment someone moves a button.

AI changes this equation. Instead of writing rigid scripts, you express intent in natural language. Click the login button. Browserbase's open source framework, Stagehand, translates that intent into browser commands. When the login button becomes a "sign in" button, the system adapts.

The use cases span from simple to sophisticated. Benny, a consumer app for food stamp users, automatically fills out rebate forms. Users take a photo of their receipt, and an AI agent navigates the government's web form on their behalf. What used to take 10 frustrating minutes now happens automatically.

KYB (Know Your Business) verification is another example. When someone opens a business bank account, the bank needs to verify information that isn't on the application form. How much funding has this startup raised? What does their website say? That's research a human used to do manually. Now an agent can check Crunchbase and Pitchbook automatically.

The model question

When asked which models perform best for browser automation, Paul gave an answer that captures the pace of AI development: "I'll say this and guarantee you in two or three weeks it won't be the right answer anymore."

Right now, Gemini is leading. Browserbase worked with Google DeepMind on their computer use model and with Microsoft on their open source Magentic-One 7B. The evals are public at stagehand.dev/evals, tracking how well different models handle real-world browser challenges like navigating complex date pickers.

The open source angle matters for enterprise adoption. Security-conscious customers want to run models in their own environments. Microsoft's Magentic-One 7B, a fine-tuned version of a Qwen model, gives them that option.

Enterprise is still early

At re:Invent, Paul attended what he called "the Davos of CIOs" meetup in Half Moon Bay. The priorities were telling: customer support ranked first for AI investment, followed by software engineering. Browser automation and back-office work? Still early.

But the trajectory is clear. Paul predicts 2026 will bring step-function improvements in computer use capabilities, with every major AI lab investing heavily in this space. The workflows that seem challenging today will become straightforward.

The metric Browserbase tracks internally says it all: years of browsing time saved. They've already deleted hundreds of years of humans going to websites and clicking buttons. Nancy at the dairy company is just the beginning.

This interview was conducted at AWS re:Invent 2025.