When Cowork launched in January, I didn't understand the hype.
We'd deployed Claude Enterprise at Nayax. Hundreds of employees now had access to a tool most of them had only dreamed about — connected to Jira, Outlook, Teams, Confluence, Snowflake. People were doing things that were simply out of reach before. That felt like the win.
Cowork existed. I knew it. But the initial version was a local file-management tool — point it at a folder, let it sort your downloads. Useful, but not what I needed. And Chat, combined with the connectors we'd spent months carefully plugging in, was already doing incredible things. So Cowork stayed on the shelf.
Then the automation requests started coming in.
The Death Spiral
Every department had something they wanted to automate. IT wanted to auto-respond to incomplete helpdesk tickets. Finance needed meeting summaries routed somewhere useful. Managers wanted Jira tickets updated based on what happened in their standups. Everyone had a list.
We started evaluating n8n and Workato. That's when the death spiral began.
How does Andy from HR run a workflow on behalf of his credentials? How many environments do we need? Who educates the entire company on workflow variables, testing, monitoring? The n8n enterprise license alone wasn't cheap. And even I — with an engineering background and hands-on software experience — had struggled with n8n. I eventually pulled off what I needed by connecting it to Claude, but when it broke, I still had to debug it in the n8n portal. That's a learning curve that shuts most people out entirely.
I didn't see a way out. No clean answer, just more governance discussions.
The Moment That Changed It
On February 24, Anthropic shipped the full enterprise push for Cowork — plugins, deep integrations, and /schedule. Boris and Thariq from Anthropic published about it. My first reaction: "What could I actually do with that?"
I sat with that question. And then it clicked — not because of the scheduling feature itself, but because of what was already underneath it.
Cowork runs in each user's context. It logs in as them. It inherits their permissions. And it connects to every SaaS integration we'd spent the previous months approving and configuring — already scoped, already governed, already ours. No new environments. No credential sharing. No governance meeting required.
The automation platform we'd been debating how to buy was already deployed.
Your Grandma Can Build the Workflow
The governance problem was solved. But I still had a nagging question: how does Andy from HR actually build his automation? He doesn't know what a "skill" is. He doesn't know how to write one.
Cowork has a skill for writing skills.
Here's how it works in practice: you have a successful conversation with Claude — doing something useful, iteratively, until it works the way you want. Then you say: "Make a skill of that." Claude understands the full workflow from the conversation, composes a skill that encodes it, then tests that skill against a clean session to see how well it performs. It produces an eval report — which steps executed, which succeeded, what needs adjusting — and refines until it's reliable.
You don't configure nodes. You don't debug variables. You speak English, do the thing once, and tell Claude to remember how.
That's when I thought: all major problems solved.
What I'm Running Today
I've been testing this since the February launch — several weeks of real use, not demos. Here's what's actually running on my account:
My work items get automatically enriched with context pulled from email threads, Teams conversations, and web research — before I even open them. When someone asks me about adding a new Claude connector, Claude handles the response: it explains the process, fetches connector details and setup instructions, and if there's no official connector, it lists relevant open source MCP servers with reputation notes. I used to write that same response manually, repeatedly.
Every few days, a usage report lands ready for IT to review license allocation. Every morning, a brief. After lunch, a lighter one. And at end of day, Claude checks what meetings I attended, summarizes them, and creates any Jira items that landed on me.
This is running. This is tested. This is not speculation.
The Next Grade
We haven't rolled this out yet. It needs its own workshop — the same way Claude Chat needed one, the same way Claude Code needed one. Each of those sessions changed how people worked. This one will too.
What we built over the past three months — the connectors, the governance, the cultural buy-in — turns out to be the foundation for something none of us fully saw coming. Not another AI feature. A personal automation layer that meets every person in the company exactly where they are.
The engineers got Claude Code. Everyone else is about to get their version of it.