Set up your AI coworker.

Who this is for

This setup suits founders, operators and small teams who lose hours each week to repeatable knowledge work: research, reports, spreadsheet clean-up, first drafts and document reviews. It works best for people who can describe the outcome they want and will check the result. It is not a good fit for tasks involving payments, legal commitments or sensitive personal data without proper controls, or for anyone who expects to hand work over and never look at it again.

Step by step setup

1. 01

   Choose a plan with enough headroom

   Cowork comes with paid Claude plans. Pro is $20 a month, or $17 a month billed annually, and suits occasional tasks. Anthropic notes that Cowork uses limits faster than chat, so daily heavy use may need Max at $100 a month. Small teams can use the Team plan at $20 per seat per month.

2. 02

   Create a dedicated working folder

   Rather than granting access to your whole documents directory, make one folder for Claude and copy in only the files a task needs. Anthropic's safety guidance recommends exactly this and advises keeping backups. Leave financial records, passwords and personal records out of it entirely. If something goes wrong, the damage is then limited to copies you can restore.

3. 03

   Connect only the tools the first task needs

   Add connectors one at a time, starting with something low risk such as a shared drive folder or calendar. Each connection widens what the agent can read and do, so add email, messaging or CRM access only when a task genuinely requires it, and remove connections you stop using. Check the permissions each connector requests before approving it.

4. 04

   Keep manual approval on at first

   Use the setting that makes Claude show its plan and wait for approval before significant actions. Move to automatic approval only for well-understood, low-stakes work. Claude always asks before permanently deleting files, but sending messages or editing shared documents deserves your attention too. Watch the first few runs of any new task closely, so you understand how it approaches the work.

5. 05

   Write tasks like a brief

   State the goal, the inputs, the output format, where to save it and what finished looks like. For example: compare these five supplier quotes, build a spreadsheet with price, delivery time and terms, flag anything unusual and save it to the Reports folder. Specific briefs produce far better results than open-ended questions.

6. 06

   Schedule only proven, low-risk tasks

   Once a task works reliably when you run it by hand, schedule it daily, weekly or monthly. Scheduled tasks run while you are away, so keep them to summaries and reports rather than anything that sends, buys or deletes, and check each run's output in the scheduled tasks list. Pause any schedule you stop relying on.

7. 07

   Review, then expand gradually

   Check outputs for errors, invented figures and missed files. Keep a short list of tasks that work well and the prompts that produced them, then extend to new areas one at a time. Treat it like onboarding a capable new colleague who still needs to learn how your business works.

What it costs

Taken from each vendor's own pricing page. Plans change, so check the link before you commit.

Common mistakes to avoid

When to get help

A general-purpose AI coworker is enough when individuals handle their own research, drafts and reports. A custom build makes sense when a workflow repeats across the team, depends on internal systems without ready-made connectors, needs audit trails and consistent output, or must run reliably without someone supervising each step. Klevere designs and builds agents for workflows like these, connected to a company's own tools and data.

Frequently asked questions

Sources

• Claude Help Center: use Claude Cowork safely