What a browser AI actually does

A browser AI sits inside your web browser and can see the page you're on — and, with your permission, act on it. Ask it to summarise the page, compare three supplier websites, pull details from a portal into a spreadsheet, or fill in a repetitive web form, and it does the clicking and typing while you watch.

For desk workers, the realistic wins are: research across multiple tabs, repetitive form entry, pulling structured information off websites, and summarising long web pages in place. If you spend chunks of your day moving information between browser windows, this category is worth understanding.

The two main options

Claude in Chrome

Anthropic's offering is a Chrome extension — it works inside the browser you already use, with the bookmarks, logins, and extensions you already have. You ask it to do something on the current page or across tabs, and it works visibly, step by step. It suits people who don't want to change browsers and who want the assistant only when summoned. It requires a paid Claude plan.

ChatGPT Atlas

OpenAI went the other way and built a whole browser with ChatGPT woven through it. Atlas is free to download, with the agent features that act on pages requiring a paid ChatGPT plan. Its strength is memory and context: it can reference your open tabs and chat history, so "draft a reply using the details from the tab with the supplier quote" works naturally. The trade-off is switching browsers — a bigger change than adding an extension.

Which one, in one paragraph

If you're settled in Chrome and already pay for Claude, the extension is the lower-friction choice. If you're happy to try a new browser and you already pay for ChatGPT, Atlas gives you the more integrated experience. Neither is worth a new subscription on its own yet for a typical office workload — try the one attached to the tool you already pay for.

The rules before you let AI drive

A browser AI acts as you — logged in as you, clicking as you. That deserves some ground rules.

Never let it work unsupervised on anything that submits, sends, or pays. Watch it fill the form; you click the final button. Keep it away from banking, payroll, and HR systems entirely. The time saved is trivial; the downside isn't. Be careful with logged-in work systems — your organisation's policy on AI tools applies to your CRM and intranet just as much as it applies to pasting text into a chatbot. If your workplace hasn't approved browser agents, assume they're not approved.

I have [number] tabs open with [e.g. supplier websites / venue options / software pricing pages]. Compare them and give me a table with: - Name - Price (note if it excludes GST) - The key inclusions - Anything unusual in their terms Don't click anything that submits a form or creates an account. If a page needs a login I haven't already done, skip it and tell me.
💡 Start read-only

Spend your first week using a browser AI for reading tasks only — summarise, compare, extract. You'll learn how often it misreads a page (it happens) before you've given it permission to act on one. Graduating from "read" to "act" after a week of accuracy is the sensible path.

Do this this week

Pick one genuine comparison task you've been putting off — three quotes to compare, several venues to assess, two software plans to weigh up. Open the tabs, run the prompt above in whichever assistant matches your existing subscription, and check the table it produces against the actual pages. You'll learn what this category does well, what it gets wrong, and whether it belongs in your weekly routine — all on a task you needed to do anyway.