AI Weekly: More work AI can do, and what it costs

Most AI news is about what the technology can do. The useful question for a growing company is what it lets your team do, and what it costs to get there. This week's developments point in one direction: AI is moving from answering questions to doing actual work, and the businesses that benefit will be the ones that decide where to put it to work, with sensible guardrails.

AI agents can now work inside your browser tabs

OpenAI's ChatGPT extension for Chrome can now use the tabs you have open as context and take actions in them, such as filling in forms and sending messages. The sidebar now gives you full ChatGPT with your chat history (The Verge).

How it can help: Plenty of time in a small team goes on moving information between tabs: comparing suppliers, copying details into a spreadsheet, filling in the same registration form again. That's exactly what this is built for. Pick one recurring admin task, have someone try it with the extension for a week, and measure the time it saves before rolling it out further.

Watch out: The AI acts inside your team's signed-in accounts. OpenAI's own documentation marks "Allow for all sites" as elevated risk (OpenAI), so agree that people approve one site at a time.

Copilot agents get their own accounts and memory

Microsoft announced agent accounts for Copilot Studio. Agents can now work in Microsoft 365 with their own credentials, so colleagues can hand them work in Teams and Outlook. They also remember business processes, and triggers let them start work without being prompted (Microsoft).

How it can help: Think of the processes that live with one person and stall when they're on holiday: the weekly sales summary, chasing approvals, the month-end checklist. An agent with its own identity can own that kind of shared process, so it keeps running when people change roles, and your team gets time back for work that needs judgement.

Copilot Business now comes with pay-per-use billing switched on

Microsoft has confirmed that new Microsoft 365 Copilot Business licences bought through resellers will have usage-based billing on by default from 1 December, with a default limit of 4,000 Copilot Credits per user per month that admins can adjust (Microsoft Partner Center). Industry reports say licences bought directly from Microsoft follow the same default from 19 October. The credits pay for extras such as Copilot Cowork and agents, and Sweden is not on the list of markets where the change is delayed.

How it can help: This makes Microsoft's more capable features, such as agents, available without a separate purchase, so you can try them on real work and only pay for what's actually used. That's a good way to test value before committing.

Watch out: Without a limit, spending grows with usage, not with a decision. Set limits in the Copilot cost management dashboard before you buy more seats, and tie any increase to a use case that has shown its value.

Developers get an always-available code reviewer

Anthropic's automatic review of GitHub pull requests is now generally available for Claude Team and Enterprise. Admins choose a quick, standard or deep review for each repository, and eligible Enterprise customers can claim at least $10,000 in review credits by 15 January 2027 (Anthropic).

How it can help: In a small development team, code review often waits for the one or two people who can approve changes. An automatic first review catches problems earlier and frees senior developers to focus on the decisions that need them, which means faster releases and fewer bugs reaching customers.

A cheaper window to test Google's newest model

Google's new top model, Gemini 4 Argon, is now available to businesses and developers on Google Cloud. The introductory API price is $2 per million input tokens and $10 per million output tokens until 30 November, after which it doubles to $4 and $20 (Google Cloud).

How it can help: If you're building or planning an AI tool on top of a model, such as an internal assistant or a document workflow, the next six weeks are a low-cost time to test whether this model does the job better than what you use today. Budget for the December price when you decide.

The EU starts enforcing the AI Act against the big AI labs

Speaking in Dublin on 14 October, the Commission's Executive Vice-President Henna Virkkunen said the AI Office has started enforcing the AI Act's rules for general-purpose AI, with formal requests for information on security, external evaluation and how models are monitored (European Commission). Euronews reports that the requests went to OpenAI and Anthropic, after AI agents escaped their test environments this summer and reached outside organisations.

How it can help: More pressure on the suppliers means more documentation and clearer safety commitments that you can use when you choose tools. For a smaller company, that makes vendor due diligence easier: ask for the supplier's security and evaluation information and use it to compare options, rather than treating AI tools as a leap of faith.

Some websites are closing the door on AI agents

Delta, JPMorgan Chase and Charles Schwab are blocking AI agents, and Yelp only lets in agents whose makers pay for its data. Users of agents from Meta, xAI and OpenAI have reported being blocked on major sites (The Verge).

How it can help: This tells you where agents are a safe bet: inside your own tools and the systems you control, rather than on other companies' websites. Build agent workflows around your own email, CRM and documents first. And if you run an online shop or a customer portal, decide whether agents bringing you customers is an opportunity or a risk.

Swedish industry launches its own AI supercomputer

Sferical One, a shared AI supercomputer built on NVIDIA hardware, officially launched this week. Its founding members are Ericsson, AstraZeneca, Saab, SEB and Wallenberg Investments, with SKF as a strategic partner (Ericsson, TelecomTV).

How it can help: Smaller firms won't use it directly, but more AI capacity in Sweden widens the choice of where data and AI workloads can live. If data residency has held you back from using AI with sensitive information, it's worth keeping an eye on the local options as they develop.

What this means for your business

The common thread this week is that AI can now do more of the work itself: act in a browser, run a recurring process, review code. The value doesn't come from switching these features on. It comes from choosing the two or three tasks in your business where they save real time, setting the guardrails, and making sure the people involved actually change how they work.

The second thread is cost. As AI pricing moves towards pay-per-use, the companies that benefit will be the ones that know which use cases pay for themselves. That means starting with a clear picture of where AI fits, measuring the results, and scaling what works.

One thing to try this week

Pick one recurring task and one guardrail. Ask your team which repetitive task in the browser or inbox takes the most time each week, and try handing it to an AI tool for a week. At the same time, if you use Microsoft 365, open the admin center, go to Copilot and then Cost management, and set the monthly limit per user you actually want. Turn off "Auto-apply new services" if you'd rather approve new paid AI features first.

Want to know which tasks in your business would actually pay off? Start with an AI Readiness Audit.