Anthropic has released a new feature called “Claude Cowork,” designed to enhance the capabilities of its AI agent, Claude, particularly for non-coding tasks. Cowork enables users to designate a specific folder for the AI to access, allowing it to read, edit, and create files. Features include the ability to reorganize files, generate spreadsheets from receipts, and draft reports from notes. Currently, Claude Cowork is available exclusively through the Claude macOS app for subscribers of the Claude Max tier, which costs between $100 and $200 per month. This feature builds on Anthropic’s earlier “Skills for Claude” initiative, which allowed users to customize Claude’s functionalities for specific tasks. The company has also highlighted potential safety concerns, noting that unclear instructions could lead to destructive actions by the AI.
Slack has announced the general availability of its updated Slackbot, which now features advanced AI reasoning, personalized responses, and the ability to utilize external tools. The updated Slackbot can summarize Slack channels, answer user questions, and draw on enterprise context without any need for training or setup. It supports integrations with third-party tools and is available in multiple languages, including English, Japanese, and French. This release enhances the capabilities of Slack as a communications platform and provides users with tailored interactions based on their access to data, including Salesforce and Google Drive. The product does not replace any previous version but expands the functionality of the existing Slackbot, making it more useful for enterprise environments.
And while I’m on chatbots, recent research from Stanford and Yale has revealed that several popular large language models, including OpenAI’s GPT and Anthropic’s Claude, have memorized and can reproduce extensive excerpts from copyrighted texts. This contradicts previous claims made by these companies, which stated that their models do not store copies of training data. For instance, Claude was able to generate nearly complete texts from well-known books like Harry Potter and the Sorcerer’s Stone and 1984, raising significant concerns about potential copyright infringement. The findings indicate that these AI models do not learn in the same way humans do; rather, they store information and output it. This phenomenon, termed “memorization,” poses a legal risk for AI companies, potentially resulting in costly copyright litigation and the need to retrain models on licensed material.
Why do we care?
There’s a quieter shift here that’s easier to miss—and harder to undo.
These tools don’t just save time. They decide what the work looks like once it’s done. When Slackbot summarizes a channel, that summary becomes “what happened.” Not because it’s perfect—but because it’s finished.
That’s the risk. AI collapses ambiguity. Humans often need ambiguity to make good decisions.
For MSPs, this drags you into advisory territory whether you like it or not. If you introduce tools that reorganize, summarize, and formalize work, you’re shaping how clients understand their business. And if no one is there to say, “Here’s what this AI simplified or ignored,” clients will assume nothing was lost.
The real harm if this is misunderstood isn’t a breach or a lawsuit. It’s slower, quieter: teams stop questioning, decisions converge too early, and AI-generated artifacts become the narrator of reality.
AI that works is valuable. AI that decides what counts is consequential. And that’s the part worth paying attention to right now.