> 📥 The full field report is available as a PDF. Download it at the top of the page — with the compliance traffic light, subscription overview and the Claude-Work folder explained.
> 🧭 Part 2 of 4 in the series AI agents anno 2026. Four tools that are all called AI agents: OpenClaw, Claude Cowork, Perplexity and Hermes. Today: Cowork — the one I use daily, and the easiest entry to a real agent.
1. Field report, not hype
I gave Cowork access to a messy project folder and asked it to clean up. A few minutes later the folder was structured, files were renamed, and nothing was lost. It would have taken me a week — and I'd never have gotten the same result. All I had to do was point at a folder and write:
"clean up this folder and create a nice folder structure that fits the content"
That was it. This is Claude Cowork. And it's something entirely different from the previous issue.
In issue 1 I wrote about my agent swarm: four agents on an old gaming PC, OpenClaw, an isolated environment, and quite a few hours of learning tax. That required technical hands. Cowork requires none. It's the easiest path into a real AI agent that exists.
2. What works
Let me start with the concrete. These are my own workflows that have survived the first weeks and still run every day.
Folder structure and cleanup. I gave Cowork access to my project folders. It reads what's there, proposes a structure and executes it. It behaves: no surprises, no accidentally deleted files.
Presentations. Cowork builds slide decks from raw notes or existing material in a folder. It matches the formatting, pulls out the right points, and delivers something 80–90% finished. The last 10–20% is my judgment. It can also apply a company logo to PowerPoint, and that works great.
Excel analyses and data enrichment. This is where it gets really interesting. Drop in a spreadsheet, and Cowork analyzes, enriches with external sources, and produces a finished report. It saves hours, not minutes. You can honestly "talk to your data" and get any chart or visualization you want. Forget building pivot tables yourself. Forget making charts in Excel or PowerBI. Sounds extreme? Try it.
The model underneath is the whole difference. Most AI tools are built around the prompt — input in, output out, next prompt. You're always in the loop. Cowork is built around the result. You describe a goal, not the next step, and it plans the path itself. It lays out a plan, shows it to you, and executes the steps with sub-agents in a locally sandboxed environment. It delivers the finished result back into your files: spreadsheets with working formulas, formatted reports, presentations. It's not copy-paste from a chat window.
Anthropic's Head of Enterprise calls it "vibe working". I'd call it Claude Code for everyone who doesn't live in a terminal.
The safety that makes it usable
Cowork asks permission. It asks before touching a new program or folder, and shows you the plan before starting. You decide what it can see. That's the whole difference between an agent you dare let in, and one you don't. You can also set it to "Act without asking" — works great when you're heading to lunch and want to make sure it doesn't stall while you eat. Cowork doesn't eat lunch, just a lot of tokens.
Dispatch and Scheduler
Cowork reaches beyond your own machine: it can take over your browser and act in webapps where you're logged in. It can talk to MCP servers and use tools like web search. On the enterprise plan there are plugins for marketing, legal risk assessments, customer support.
Dispatch is the function I've tested most thoroughly — and the one that has disappointed me most. The idea is beautiful: throw a task from your phone, let the desktop execute while you're in a meeting. In practice it's limited. You can't ask it to continue a project on the go unless you set it up specifically from your Mac first. On top of that, there's a hard technical limit: the Mac has to be open and running.
On the other hand, one function has surprised me positively: Scheduler. It's phenomenal. It runs flawlessly and delivers good news to my Gmail draft folder every morning. So far the most direct competitor I've seen to the OpenClaw setup from issue 1 — just built in and controlled via a long prompt Opus set up for me.
And that's where Cowork's real strength lies. In issue 1 you had to be technically strong even to get started. Cowork runs as an app on Mac or Windows, sits on the paid Claude plans from $20/month, and a non-technical colleague is up and running in minutes.
3. What it can't do
Here's the honest part. After weeks of daily use, my assessment: Cowork performs like a PhD student with nearly endless memory. It has read everything in the folder. It connects, synthesizes and produces at a pace that's hard to match.
And then it makes mistakes a 6-year-old would make. It has misspelled words in headlines. It once invented a brand new surname for me and put it on the front of a PowerPoint. Output that's 95% perfect and 5% embarrassing.
The consequence is simple: you have to check everything. Always. It's the price, and it's fair — because what you get back is still hours saved every week. But an agent you don't proofread is a risk, not a help. The agent does the work, the human owns the judgment — just on a tool where it's even easier to forget.
4. Cowork touches your real files
Cowork gets access to your folders, your files, and your programs. It can read and write real documents. That shifts the whole compliance question.
With a chat, it's primarily about what you type in the field. With an agent, it's about what the agent has access to. That distinction matters more for some industries than others.
In a compliance assessment I did for an audit firm, we ended up with a simple traffic light:
| Data type | Status | Prerequisite |
|---|---|---|
| Internal working documents without personal data | GREEN | Default setup is typically sufficient. |
| Client deliverables with ordinary personal data | YELLOW | Only possible with data minimization. Legal basis and human control of the result are required. |
| Sensitive personal data + full KYC folders | RED | Compliance will often say no without a separate risk assessment, access restrictions and documented audit trail. |
Compliance differs depending on the Cowork subscription. On a private Free, Pro or Max plan, consumer terms apply. Chats and coding sessions can be used in model training if "Help improve Claude" is on, and conversations can be analyzed by Anthropic on safety flags. That's red for client data, confidential business data, source code and personal data.
Browser and computer control deserves a separate word. On paper it's strong. From a compliance perspective it's also the riskiest part. I've experimented with it, but I don't let it buy anything for me, and I don't give it access to places with two-factor: bank, email, accounting. As a standard tool, it's not there yet.
5. Who it's for
If a non-technical colleague needs to get started with a real agent this week, Cowork is hard to beat. Low threshold, no setup, fast value. For teams that want to give their people an agent without an IT project, it's the strongest bet right now. Buy an Enterprise/Teams plan and add colleagues. You're up and running in 10 minutes.
Microsoft houses: watch Copilot Cowork
Microsoft has its own version, Copilot Cowork, built on the Claude models. It's the most promising path to an enterprise-mature version: one vendor, familiar admin controls. But for European companies, Claude in Microsoft 365 is currently outside the EU Data Boundary and disabled by default for EU tenants. The data residency issue isn't solved yet. But it's the development I'd watch closest.
6. How to get started
Cowork is the opposite of my agent swarm from issue 1. Where OpenClaw required technical hands and an isolated environment, Cowork requires a folder and a sentence. It's the lowest threshold of all four tools in this series.
But precisely because the threshold is so low, it's easy to skip the important question. Cowork touches your real files. Choose the subscription deliberately, choose the folders deliberately, and proofread everything.
And there's one thing I wish I'd known from day one: the central Claude-Work folder.
| Element | Description |
|---|---|
| Claude-Work/ | One central hub folder on your machine. The home base Cowork points to first. |
| Projects/ | Each project as a subfolder — or pointing to where the real files live. The actual work happens out there, not in the hub. |
| CLAUDE.md | A short text file at the top of the folder telling Cowork what the folder is, who you are, what you do, what your voice is, and how the folders are organized. Two-three paragraphs is enough. It reads it every time. Anthropic recommends up to 200 words. Don't invent it yourself — have an Opus 4.8 high build it for you (remember meta-prompting!). |
Once the structure is in place, you stop having to explain everything from scratch every time. Cowork already knows who you are, what your voice is, which templates you use, and what a good output looks like. The result becomes dramatically more "you".
A few practical tips:
1. Start with the Claude-Work folder. Don't point Cowork at your entire Desktop or Documents library on day one. Give it a bounded, controlled base.
2. Put your best deliverables in. Templates, examples, "this is how I usually write".
3. Choose your first real task carefully. Cleanup, structure, or a slide deck built from existing notes. Low-risk, fast result.
4. Keep client data out of the folder for now. Give it access to client data, personal data and confidential documents when you know how the tool behaves, and when the subscription is the right one.
5. Always read the result through. Always. The 5% error is always waiting somewhere.
Next issue: Perplexity Computer / Research. Not a general agent, but a specialist. A research flow that takes you all the way from question to finished document. What it can do, where it's strong, and why "impressive" isn't the same as "GDPR-ready".
Stefano Vincenti — GenAI strategist and architect. Co-founder of BotTellMe. External lecturer at ITU and DIS Copenhagen. Partner at TryZone. Subscribe to the newsletter and get the next two issues directly.