AI agents have moved past the demo stage for businesses. Claude can now browse a website just like a customer would. I’ve been testing that new agent and other Claude features (like the new M365 write access and a recording skill) inside our managed IT services company for weeks. They’re great, but there are a few catches worth knowing before your team uses them. So, let’s start with the most exciting feature first: Claude’s new browser agent.
Key Takeaways
- Claude’s Chrome extension can research and act inside a browser on its own, pulling information from more than one website without anyone clicking through by hand.
- Claude plugins bundle skills that are already built, plus the data connectors, into toolkits built for one job, like finance or operations.
- Claude’s Microsoft 365 connector gained tools to create, edit, and delete data in 2026.
- Recording a skill lets someone teach Claude a task by demonstrating it once.
Table of Contents
How Claude's Browser Extension Turns AI into an Agent That Acts
Most people who use Claude stick to the desktop app or the website. Few know about the Claude in Chrome extension, which sits in the browser toolbar next to your password manager or ad blocker. It currently only runs in Google Chrome, and it needs the desktop app installed on the same machine to actually work in the browser.
What can Claude do once it's in your browser?
When it’s opened on any page, it can read that page and answer questions you have about it. That is a small time saver. But the more useful part is what happens when Claude is given a task instead of a question. It can open tabs, click through a website, and pull information back into a conversation, without you having to do any of the clicking.
The really important thing with any AI tool is the context. What is the information that it is looking at when you’re prompting it. It’s the same reason a new hire needs an orientation before they’re useful on the job. They can’t act on what they don’t know yet. The browser extension hands Claude that context automatically, reading whatever page someone’s already looking at instead of making them type out a description first.
What does this look like in actual work?
I recently pulled usage statistics out of the Microsoft Partner Center for a group of clients using Microsoft Copilot. Normally, I would have had to go click, click, back and forth, click, copy, paste, click, click, click. But instead, all I had to do was prompt Claude to use the Chrome extension. It opened the Partner Center pages, read through them, and brought the data back into my conversation on its own.
The same approach also works for less technical tasks. For example, in the video I ask Claude to visit tealtech.com’s services pages and Teal’s CMMC compliance site and summarize what it found.
Claude opened a browser window, clicked through each page, and read the content, while I kept working on something else. This is a really powerful business automation tool now. We’re actually asking it to use my computer to use the browser to go out like a human would, and read websites, and interact with them. Each new site it visits asks for permission first, so someone has to be present to approve those initial requests.
What Claude Plugins and Connectors Can Do for Your Business
Two features get less attention than they should, mostly because they sound similar and live in the same settings menu. Those are Claude plugins and Claude connectors. A connector is a link between Claude and an outside system, like Google Drive, Notion, or Microsoft 365. These connectors let Claude read (and now sometimes write) data.
A skill, on the other hand, is a set of written instructions that Claude follows to complete a specific task. A plugin bundles both together. So, you get a set of skills built for one job function, like finance or operations, plus the connectors those skills need to do their job.
A finance plugin, for example, might include a skill for reconciling a spreadsheet, ready to run the moment someone types a slash command. Third parties build plugins too. Apollo has published skills for enriching leads, building prospect lists, and loading outreach sequences, on top of the Apollo connector many sales teams already have turned on without using the packaged skills.
Installing a plugin is easy and takes just a single click.
The catch is that every connector you add forces you to consider the data that Claude can reach. A connector like Zapier, which links to hundreds of other tools, can open a business up to systems nobody on the IT side can see or approve. So, a user that connects Claude to their private Zapier account can create risk that no one in the business knows exists.
So, before anyone installs a plugin or turns on a connector, someone in IT should know what data it touches and whether the business should trust the organization that built it.
What Changed When Claude's Microsoft 365 Connector Added Write Access?
Until mid-2026, Claude’s Microsoft 365 connector could only read data. So, it could simply search emails, browse a calendar, or open files. But in July 2026, Anthropic added tools that let Claude create, edit, and delete data in Microsoft 365 too. That means you can now make a request like “Add this to my calendar,” and it’ll actually happen.
The connector originally defaulted to always allowing reads, and our team left that in place. But before turning on any of the 30 new write and delete tools Anthropic released, we configured the connector to require approval first. So, whenever you’re changing your data, deleting your data, you want to make sure that you’re getting approval. It’s important to keep the human in the loop before Claude changes anything.
Be sure to set up approval and review before turning features on. The NIST AI Risk Management Framework treats this kind of human review as a standard control for higher-risk AI actions.
It’s important to mention that the M365 write and delete tools are turned off by default. So, you’ll need to deliberately turn each one on.
How does Claude's Microsoft 365 writing stack up against Copilot, and why does adoption timing matter?
Claude’s writing tools don’t necessarily beat Microsoft’s own Copilot writing capabilities yet. I haven’t personally used it enough yet to say one or the other is better. What I have noticed is that Claude handles large volumes of Microsoft data well, working through months of calendar history the way Copilot would.
The timing matters too. Gartner predicts that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today, a pace few other technologies have matched. At the same time, Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls.
While this seems contradictory, both predictions can be true at once. The technology is simply moving very fast, and many business leaders aren’t sure how to get the max value from them.
The truth is, the businesses that will see the most value from AI and automation will be the ones that:
- Set up the tools to improve their workflows
- Drive adoption through monthly training
- Create an acceptable use policy
- Build a defense-in-depth security foundation + additional layers to protect the AI model itself and how their team uses it
How Recording a Skill Turns an Action into a Repeatable Workflow in Claude
The last feature we’re going to look at is the newest, and the most limited in where it works because it only works in the desktop app, and only on a Mac. Inside the Cowork mode of Claude’s desktop app, a “record a skill” option watches someone’s screen and listens to them narrate while they perform a task. Claude turns that recording into a written set of steps, which can be reviewed, edited, and saved as a reusable skill.
What's the security risk when recording a skill, and how do you avoid it?
It’s a very powerful feature. That being said, you want to make sure that you never type a password while you’re recording, for obvious safety reasons. Skills are just text files once they’re created, and they can be shared with a team the same way a document can. So, anything typed during the recording, including credentials, can end up written into that file. I recommend that you sign into any site you plan to use first, then start recording. That way, nothing sensitive will ever make it into the instructions Claude writes for you.
What does this look like when a team uses it?
I’ve used this capability to record a skill against our company’s ticketing system, which runs in a browser. I log in, search for a customer’s tickets, and download the results, once, while Claude watches/records. After that, a request like “Get me the tickets for this client” will pull up the data using the skill, without repeating all the clicks by hand.
Another use case might be pulling the numbers our team reviews every week for a leadership meeting, which today means visiting several different websites by hand. We simply use the skill and it does gets everything we need much faster than it would take us to complete manually.
What if I don't have a Mac?
For anyone without a Mac, there’s a slower path to achieving the same result. You just need to work through a task with Claude step by step verbally. Confirm that each part/step works the way you intended, and then ask Claude to turn the entire conversation into a skill. It takes a bit longer to build, but it produces the same kind of reusable instructions to save you time in the future.
What Business Leaders Should Ask
None of these features require a deep technical background to understand, but they all require a decision about how much autonomy to hand over, and to whom.
We started small. We have the minimum Claude Team account, with five seats, and we have since added only one or two more. Those seats are used constantly in production work. That kind of narrow, deliberate rollout is worth copying.
Before enabling a browser extension, a plugin, a connector, or a recorded skill, there are four questions you should ask:
- Does this feature need write access to change or delete data, or is read-only enough for what the business is trying to do?
- Who has to approve before Claude changes or deletes something in a live system?
- Which connectors are already turned on across the business, and does IT know about every one of them?
- How many people actually need access, versus how many currently have it?
Where AI and Agents Go from Here for Businesses
AI agents are already in the browser, the inbox, and calendar of anyone running Claude with these features turned on. The technology works well for businesses that decide upfront which actions need a person to review before they happen.
Claude also has a computer use capability now – which is really powerful. But for that to be valuable, we have to be thinking about what the processes are that we do every day, every week, and every month. Most businesses haven’t done that mapping work yet. That’s a significant bottleneck now, not the technology itself.
Our AI consulting and automation engineering services are perfect for companies without the IT capacity to map that work and configure the right guardrails. Our team can build the security you need, provide training, and help you build the workflows that will grow your business. Reach out if you’re interested in speaking with me about how we can help.
