Your First 30 Days of MCP: A Pilot Rollout Playbook

An MCP pilot is a short, controlled trial where one team connects an AI app to a few business systems and uses it for real work. It is usually led by an operations, IT or AI lead with a sponsor from the business side. MCP changes rollout planning because each connection gives an AI app live access to a system, so the pilot has to test access controls as well as usefulness.

This playbook lays out the first 30 days, week by week. It uses admin controls that Anthropic and OpenAI document for Claude and ChatGPT as of October 2026.

What you need

• A business AI client with admin controls. On Claude Team and Enterprise, an Owner adds a connector for the organization and members then connect with their own accounts, per Claude's connector docs. On ChatGPT, OpenAI's custom MCP server article says custom MCP connections are available for Business and Enterprise/Edu on the web.
• Two or three official MCP servers for the chosen workflow. Use vendor-published servers where they exist; our server-selection checklist explains how to judge them.
• A pilot group. Anthropic's Claude Enterprise administrator guide suggests a broader Claude pilot of 50 to 100 users. For an MCP-specific pilot, a single team is easier to support.
• A log source, from the AI client or a gateway, so you can see which tools were called.
• A one-page usage rule that says what staff may and may not connect.

The workflow

Week 1: choose and prepare

1. Pick one workflow that is frequent, low-risk and easy to check, such as summarizing support tickets or drafting weekly status notes. Write down how it works today and roughly how long it takes.
2. Pick the servers. Choose only the servers the workflow needs. Record the owner, the OAuth scopes requested and which tools can write or delete.
3. Set up admin controls. In Claude, the Owner adds the connectors from Organization settings > Connectors. In ChatGPT, OpenAI says admins and owners can "connect and test MCP servers privately before publishing" them for the workspace, and only admins and owners can publish.
4. Set approvals. In Claude, each connector has Tool permissions with Always allow, Needs approval or Blocked. ChatGPT's app permissions run from Always ask to Allow all actions, per OpenAI's permissions article. Start with approval required for anything that writes.

Week 2: launch to the pilot team

1. Run a 45-minute kickoff. Show the workflow, the approval prompts and what to do if a tool acts strangely. Our staff training playbook has an outline.
2. Name two or three champions. Anthropic's guide recommends "2–3 champions per department or team" with early access and extra training.
3. Open a feedback channel. Ask for examples: a good result, a bad result, and any approval prompt that confused someone.

Week 3: observe and adjust

1. Review the logs weekly. Check which tools were called, by whom, and whether any writes happened without the expected approval.
2. Tune permissions. Block tools nobody needs. Move a tool from Needs approval to Always allow only if it is read-only and the team has approved it many times without issue.
3. Fix the instructions. Update prompts and the usage rule based on what went wrong.

Week 4: decide

1. Compare against the baseline you wrote in week 1, using the measures below.
2. Write the decision. Go (expand to another team), adjust (run another month with changes) or stop. List the servers you would approve permanently; they become the start of your approved server list.

Example prompts

Adapt these to your chosen servers:

• "Using the support connector, list the tickets opened yesterday that mention billing. Summarize each in one line. Do not reply to any ticket."
• "From the project tracker, list tasks due this week that have no assignee. Show them in a table."
• "Draft a status note from these three tasks. Show me the draft before you post anything."
• "Which tools does this connector give you? Which of them can change data?"

Guardrails

• Read-only first. The MCP tools specification says there "SHOULD always be a human in the loop with the ability to deny tool invocations." Keep that true for writes during the pilot.
• No personal connectors on company data. Stick to the connectors the admin added.
• Keep sensitive systems out. Leave payroll, HR files and payment systems out of a first pilot.
• Plan for misbehavior. Agree who can turn a connector off, and how fast. See the MCP security center for the risks to brief your team on.

What to measure

• Time spent on the workflow before and after, as reported by the team or timed on a sample.
• The share of AI results the team accepted without major edits.
• Tool calls per user per week, from logs.
• Approval prompts shown, accepted and denied.
• Incidents and near misses, with a short note on each.
• Pilot users' answer to "Would you keep using this?"

Do not borrow results from other companies. Your own before-and-after numbers are the evidence. Our ROI playbook covers how to set up the comparison.

Next step

If you want a second opinion on which workflow to pilot and which servers are safe to start with, our MCP readiness audit covers both. Smaller teams can start with MCP for small business, and more guides are in the MCP playbooks hub.

Frequently asked questions