Roll out Claude one workflow at a time, not one license at a time.

Where should a Claude rollout start?

Start with one workflow that repeats every week, has a clear owner and produces something you can check. A company-wide launch with no named use case produces a burst of curiosity and then quiet.

Good first workflows are ordinary: customer follow-up after a call, a weekly report, first-pass review of incoming documents, proposal drafts from a template. For each candidate, write down who does it, how often, what goes in, what comes out, and who checks it. Pick the one where the output is easiest to judge. If you want a structured way to do this, a workflow audit is exactly that exercise.

Name two to four people who own that workflow. They are the pilot. Everyone else waits until the pilot has something that works.

Which Claude plan does a team need, Team or Enterprise?

Team covers most small and mid-sized businesses. Enterprise adds identity provisioning, audit and compliance controls that larger or regulated organizations need. The split below is how Anthropic's plan page described the two in September 2026; plans change, so confirm on the current page before you buy.

The practical questions: do you need to provision and remove accounts from an identity system automatically, do you need an audit trail or spend limits per person, and do you handle health information? If all three answers are no, start on Team. Avoid individual consumer accounts for company work; they sit outside your admin controls.

How should Projects and Skills be set up?

Give the pilot workflow its own shared Project, then turn the steps that keep working into a Skill. Projects hold context for a body of work. Skills hold a procedure Claude follows whenever the task comes up.

1. Create one Project per workflow, not per person. Name it after the work, such as "Client follow-up".
2. Write the Project instructions: who the output is for, the house tone, the format, and what Claude must not invent. Keep it to what actually changes the output.
3. Add knowledge files the task depends on: the template, the price sheet, two good past examples. Remove anything stale.
4. Share it with the pilot group and have everyone run real work through it for two weeks.
5. Promote what works into a Skill. When the same sequence of steps is being typed out again and again, package it as a Skill so it loads automatically when the task matches. Team and Enterprise admins can deploy Skills across the organization. How to write a SKILL.md that actually gets used covers the part people get wrong, which is the description that makes it trigger.

Which connectors should you turn on, and where does Cowork fit?

Turn on only the connectors the pilot workflow needs. Every connector is access to company data, so treat each one as a decision, not a default.

Connectors let Claude reach systems such as email, calendar, file storage or a CRM. On Team and Enterprise, Owners control which connectors are available, and only an Owner can add a custom connector built on a remote MCP server. Each person then connects and signs in individually, so Claude sees only what that person could already see. If a system your workflow depends on has no connector, that is the point where a custom one is worth scoping.

Claude Cowork, in the Claude desktop app for macOS and Windows, is for work that lives in files. Claude reads and writes inside a folder you approve, so a spreadsheet or memo is built in place rather than uploaded and downloaded. Give it the project folder it needs, not the whole machine. Owners switch Cowork on or off for the organization under Organization settings, and as of September 2026 Team plans also have cloud sessions, which run tasks on Anthropic's infrastructure, turned on by default. Decide whether you want that before the pilot starts. On September 16, 2026, Anthropic announced that Cowork and chat are merging into one Claude app, rolling out to Pro and Max first, then Team and Free, with at least 30 days' notice for Enterprise; existing chats, projects, connectors and Skills carry over. Expect the Cowork name to disappear from the interface while the folder access and admin settings stay.

What belongs in the data policy?

A page or two that answers what people may paste in, what needs permission, what is off limits and who reviews output before it leaves the building. Write it before training, so training can teach it.

• Approved tools and accounts. Company work happens in the company workspace.
• Data tiers: fine to use, needs a manager's approval, never. Name the categories: customer records, employee files, financial data, contracts, credentials, anything regulated.
• Review rules: anything customer-facing, financial or legal gets a human check before it is sent.
• Connector approval: who decides a new connector is allowed, and who owns it.
• Personal AI agents: none connected to company email, calendars or files unless approved. Consumer agents now ask for exactly that access.
• What to do when something goes wrong, and who to tell.

Our AI governance lite guide has a fuller template sized for small teams.

How should the team be trained?

Train on the pilot workflow with the team's own files, not on a feature tour. People keep what they practiced on their own work and forget what they watched.

A good session has each person run a real task end to end, compare the output with what they would have written, fix the instructions, and leave with the Project and prompts already working. Teach judgment alongside mechanics: what to verify, when to rewrite and when a task should not have been handed over. altr runs this as a paid enablement engagement, either a working session with the one to four people who own a process or a team workshop for up to ten. More on what makes training stick is in AI training for employees.

How do you know the rollout worked?

Measure the workflow, not the logins. Seat activity tells you people opened Claude. It does not tell you the work changed.

• Use on the target task: how many instances of the workflow ran through the Project or Skill this week.
• Time per instance: a rough before and after, timed by the people doing it.
• Rework: how often the reviewer sends output back, and why.
• Errors caught in review: each one is a fix to the instructions.
• Owner's verdict: would they go back to the old way?

When the first workflow holds up for a month, pick the second one and repeat. The Project, the Skill and the policy you already have make the next one faster.