Anthropic recently released a guide for rolling out Claude Cowork across an organisation. Cowork is a step beyond the chat window. It’s an agent that works directly in the employee’s local files, Slack, Google Drive and browser. It integrates with Excel and PowerPoint, and can be extended with plugins, skills and custom commands.
For many organisations this will be the first real AI rollout. And with that, the first real test of AI governance.
Because what Cowork does in practice is move the decision about AI use from the IT department to every employee’s desk. Shadow AI becomes mainstream AI overnight. And what used to be a future question for the leadership team is suddenly an operational one for every team.
The tool isn’t what’s new. The reach is.
Claude in a chat window is a contained experience. The user pastes in text, gets an answer, copies it back. The data moves between two clear points. The risk can be reasoned about.
Cowork removes those boundaries. The agent has access to local files. It reads and writes in Slack conversations. It retrieves and processes documents in Drive. It fills in spreadsheets and builds presentations. It acts in the browser.
This isn’t an extension of a chatbot. It’s a new category of work tool. And that’s why Anthropic themselves describe the rollout as a five-stage maturity ladder rather than an installation.
Three questions that decide whether your pilot holds
Before the first team opens Cowork, the leadership team needs to be able to answer three questions. Not in the abstract. Concretely, documented, communicated.
Which decisions may the agent make without human review? If an analyst asks the agent to “send this to the client”, is that mandate enough? If HR asks it to “update the salary file”, who owns that outcome? As we noted in The maturity report is written for the board: a board that hasn’t rehearsed the decision will be making it for the first time under pressure.
Which data may the agent touch? An agent with read access to local files will read every file it has access to. The data catalogue from 2019, the old project on Drive, the export from the HR system nobody cleaned up. What happens when that ends up in a summary? And where does the summary go?
What happens when something goes wrong? Not if. When. Who notices? Who reports it? How do you tell apart an agent that misunderstood the task, a user who misused the tool, and an actual incident that needs escalating to the regulator?
These aren’t hypothetical questions. They’re the questions every organisation covered by NIS2, the Swedish Cybersecurity Act or the EU AI Act will have to answer at the first review. The difference is whether the answer exists before or after an incident.
Anthropic’s maturity ladder is sound. But it isn’t complete.
Anthropic proposes a rollout in five levels: from simple chat Q&A, through team-specific integrations, to organisation-wide plugins and skills. It’s a good structure. It reflects how AI use actually matures in an organisation, from experiment to operational capability.
But the ladder describes only one half of that maturity.
For every step up in AI maturity there needs to be a corresponding step up in security maturity. Otherwise you’re building speed, not capability. An organisation at AI level three with security at level one hasn’t rolled out AI. It has deferred the problem.
That’s why in AI FastTrack we work with two ladders in parallel. AI maturity and security maturity move in step, one rung at a time. No AI pilot decision is taken without the corresponding control points in place. No step up without the underlying governance being able to carry the load.
It isn’t a brake. It’s what lets you keep going.
Champion teams are the right idea. But not enough on their own.
Anthropic recommends starting with a champion team that leads the way. That’s pragmatic and proven. It works.
But it assumes the champion team already has access to a governance structure the rest of the organisation can inherit. Roles are defined. Decision rights are clear. Data classification is done. Logging is in place. Escalation paths have been tested.
Without that infrastructure, the champion team doesn’t become a pioneer. It becomes a proof of concept for exactly what shouldn’t be scaled: AI use without formal governance. When other teams follow, they don’t inherit structure. They inherit a vacuum.
The most common misconception we meet is that governance can be shaped during the rollout. It can’t. Governance has to exist before the pilot, not after. Otherwise it’s no longer a pilot. It’s a rollout nobody actually approved.
What you can start on tomorrow
There are things you can begin with straight away, without waiting for a formal method. Three examples.
Take stock of which AI is already being used in the organisation, and how. That picture is almost always larger than leadership thinks.
Identify the three to five decisions where an agent would make the biggest difference. Not the most spectacular ones. The most repetitive ones.
Decide where you don’t want an agent operating, and write it down. Drawing the lines up front is usually easier than finding them afterwards.
It’s a start. It doesn’t get you all the way.
When it gets serious
AI FastTrack is our seven-week method for leadership teams that want to turn AI ambitions into governable business impact. We start with decision mapping, not tool selection. We map security maturity against AI maturity at every step. And we leave you with a concrete roadmap and a maturity ladder your board can follow up on, whichever AI vendor you choose.
Cowork may be new. The questions it raises are not. They’re the same questions every other tool with desktop access has raised: who owns the decision, who sees the data, what happens when something goes sideways. The difference is that the answers need to be in place before people start using the tool, not after.
Because once AI is on the desktop, it’s too late to work out where the decisions should be made. By then they’re already being made, a hundred times a day, by agents that don’t know what they don’t know.
Do you need help building governance that matches your AI ambition? Contact us for a free consultation, or read more about AI FastTrack.
More insights
Related articles
The quantum threat is a leadership question, not a technical one
The quantum threat isn't about when the quantum computer arrives, but how long your information must stay secret. Here's how to begin the move to post-quantum cryptography.
Our own software won't solve your problem
We build and sell software for security and compliance. So this might sound strange: the software won't solve your problem. Not by itself. Let us explain what it actually can't do.
The maturity report is written for the board. Not for the crisis.
Status reports test no one. When the alarm goes off at 02:14 on a Friday, what matters is the board's ability to decide, not the maturity score. Practise, don't just report.