AI browser agents: what they can already do for your business

On August 26, 2026, Anthropic announced that Claude in Chrome — the extension that lets the model browse, click and fill in forms on its own — is now generally available on every paid plan, out of its earlier pilot phase. It confirms something that had been building for weeks: earlier in the month, Claude Cowork got its own built-in browser inside the desktop app, no extension required. For any SME already using AI to draft or summarize text, this is a real shift: from an assistant that answers questions to an agent that carries out whole web tasks while the team works on something else.

What changed this week

Until now, this kind of autonomous browsing required approving every single action — every click, every field filled in — which made it fine for testing but impractical for daily use. With the August 26 update, Claude can now chain actions together without asking for confirmation at every step, as long as a safety classifier first validates that the action matches what was asked. In practice: reading a page, filling in a form, moving to the next tab and finishing a multi-step task without anyone watching the screen in real time (Anthropic, August 26, 2026). Independent coverage confirms the same facts: the move from pilot to general availability on every paid plan, with the safety classifier as the centerpiece of the announcement (Gigazine, August 27, 2026).

The missing piece arrived a few weeks earlier: Claude Cowork started opening a built-in Chromium browser inside its desktop app (Mac, Windows and Linux) whenever a task needs one, without relying on the user’s own browser or any extension. That browser can’t see the team’s tabs, passwords or bookmarks, asks for permission the first time it acts on a new site, and blocks high-risk sites by default.

What a browser agent can already automate

For any business that runs part of its operations through a browser — which is nearly all of them — this opens up very concrete tasks:

• Repetitive research: comparing competitor prices, finding suppliers that meet a set of criteria, pulling public data from several sites into one table.
• Forms and online paperwork: placing orders on supplier portals, filling in tender or registration forms a person currently handles by hand.
• Periodic checks: watching whether a competitor’s page has changed price or stock, or whether a public procedure has changed status, without anyone having to remember to check.
• Multi-step chained tasks: finding a figure on a website, copying it into a spreadsheet and sending a summary, all from a single instruction.

None of these tasks is new by itself; what’s new is that they no longer need custom development to get started — just a natural-language instruction and a paid account.

The limits worth knowing before granting access

An agent that browses public sites with non-sensitive data is reasonably safe today. Things change once a system holds customer data, access credentials, or actions that cost money — payments, placing orders, changes to a CRM. That’s where it pays to go slowly:

• Team and Enterprise plans let an admin configure allowlists or blocklists of sites, and disable the extension organization-wide if needed. On an individual Pro plan, that control doesn’t exist: each person decides what access to grant.
• Safety classifiers against prompt injection — hidden content on a page trying to make the agent run actions other than the ones requested — reduce the risk, but don’t remove it. Before automating anything with financial or legal consequences, it’s worth testing it in a controlled setting first.
• A generic browser agent works well for one-off, low-risk tasks. For something that repeats every day and touches internal systems — a CRM, an ERP, billing — a custom-built agent connected directly to those systems is usually more reliable and cheaper to maintain long-term than chaining instructions to a browser agent each time.

When off-the-shelf is enough, and when it’s worth building something custom

The question we hear most often at Evicron, an AI and custom software studio based in Barcelona with 200+ projects delivered across 12 industries since 2019, isn’t whether AI can do a task, but whether it’s worth solving with an off-the-shelf tool or a custom build. The answer comes down to volume: if a task happens a few times a week and tolerates human review, a browser agent like the ones described here is the fastest, cheapest option. If it happens dozens of times a day, touches sensitive data, or needs to reliably integrate with an internal system, it’s worth investing in something custom-built, as we cover in more detail in our 2026 applied AI pricing guide.

Bottom line

As of August 26, 2026, browsing and completing web tasks autonomously is no longer a beta feature — it’s available on every paid Claude plan. It’s already a useful tool for research, forms and periodic checks, with reasonable safety controls for both individual and team use. For anything touching internal systems or repeating many times a day, it still pays to weigh a custom build before automating it with a generic agent.

Want to know whether your business can automate a specific task with off-the-shelf AI or needs something custom? At AI consulting we look at your case through a free discovery call. Get in touch and we’ll give you a clear answer within 24 hours.