You've been asked to set up Claude for the company. Here's where to start.

You probably did not apply for the role of "company AI lead." It got handed to you — by a founder who doesn't want to deal with it, a department head who doesn't know where to start, or a CEO who read something about AI and sent you a Slack message at 9pm.

However you got here, you are now the person. Here is what the first two weeks actually look like.

Week one: understand before you configure

The biggest mistake new admins make is configuring things before they understand what the team actually does. They set up Projects, write system prompts, enable Skills — and then discover three weeks later that the CS team's most common task is not ticket responses, it is internal escalation notes.

Spend week one talking to people. Not asking "what do you want from AI" (you will get fantasy answers). Ask instead:

• "What do you do most of that is repetitive and takes longer than it should?"
• "What kind of writing do you produce in a typical week?"
• "What information do you spend time looking up or gathering?"
• "What have you already tried with AI, and what happened?"

That last question is the most important. Most teams have already experimented informally — some people are using Claude on their personal accounts, others tried it and gave up. Understanding what they tried and why it worked or didn't is the fastest path to knowing what configuration will actually help.

The four things to figure out in week one

1. Which teams have the clearest use cases?
Look for teams with high-volume, repetitive, text-heavy work: customer success, marketing, operations, HR. These are where Claude delivers fastest. Engineering is valuable but complex. Sales is high-value but needs more setup. Start with clarity.

2. Who will be your Project owners?
You cannot write every team's system prompt — you don't know their work well enough. Identify one person per team who understands the work AND cares about quality. That person becomes the Project owner. Your job is to give them a framework and stay out of their way.

3. What plan do you actually need?
If you are deploying to 5+ people, you need the Team plan — not individual Pro accounts. The admin console alone is worth it. See the full plan comparison in Which Claude plan is right for your organisation?

4. What is off-limits?
Before anyone sends a customer email drafted by Claude, decide: what outputs require human review? For most organisations, customer-facing content, anything involving personal data, and legal or financial commitments should have a human in the loop. Write this down. Tell the team before they start, not after someone sends something they should not have.

Week two: build the first Projects

Once you know what the priority teams need, build their Projects. See the org-wide architecture guide for the full structure. The short version:

1. Name Projects clearly: Customer Success — Ticket Drafts, not "CS Team"
2. Write the system prompt with the Project owner, not for them
3. Upload 3–5 documents per Project, not 50
4. Enable the Skills each team needs; disable the ones they don't
5. Test with 10 representative tasks before anyone else touches it

Roll out to 1–2 early adopters per team first. Let them break things and report back. The edge cases they find in the first week are the most valuable input you will get.

The thing nobody tells you about this job

Your job is not to make Claude work. Claude already works. Your job is to make Claude useful for your specific team, with your specific context, doing your specific work.

That requires understanding the work. The more time you spend understanding what each team actually does before you configure anything, the better the outputs will be. Admins who skip this step spend the next six months fixing a system that was built on assumptions.

The teams that get the most out of Claude are not the ones with the most technically sophisticated setup. They are the ones where someone took the time to understand the work before writing the first system prompt.

That someone is you.

What good looks like at 30 days

At one month in, you should have:

• 3–5 active Projects with real owners
• At least one team that has noticeably changed how they work
• A list of what is working, what isn't, and what to fix
• A plan for the next wave of teams to onboard

You should not have: 20 Projects, a comprehensive usage policy nobody has read, or a weekly AI newsletter you are already behind on writing.

Do less, better. That is the job.