How can your team manage shared AI workflows?

When several people share an AI workflow

A campaign team might use a shared AI project to draft updates, check research, or prepare supporter replies. As staff and volunteers join the work, it helps to agree who can change the instructions, approve outputs, and manage access.

In tools such as Claude and ChatGPT, project membership, access to connected files, and permission to take actions can be separate controls. The options below can help your team decide which responsibilities and checks fit its work.

Three permission questions to consider

Adding someone to an AI project answers only one access question. A useful review separates three controls:

1. Project access: Who can find, use, edit, share, archive, or delete the AI project?
2. Source access: Which files, folders, records, and connected systems can each person or service account open?
3. Action access: What may the AI read, draft, create, change, send, publish, or delete?

An AI connector usually works with the permissions of the account that authenticated it. It does not repair an over-shared Drive, CRM, or mailbox. Reviewing source permissions before connecting an account helps catch material the AI workflow should not be able to reach.

Individual accounts make it easier to trace activity and remove one person’s access than shared logins. Organization-managed groups can simplify onboarding and offboarding by changing access in one place.

Roles your team can adapt

These are responsibilities your team can assign, rather than a list of settings available in every AI tool. One person may cover several; your team may not need all of them.

Some products bundle file editing, instruction editing, and member invitations into one role. For someone who only needs to chat or draft, a use or view role can avoid unintended changes to shared context. People maintaining that context need broader access, including any extra permissions the product bundles with it.

Visibility also differs by product. In a shared ChatGPT project, members can view the project’s chats and files. In a shared Claude project, members can use shared knowledge and instructions, while chats remain private unless shared. This distinction matters when deciding what belongs in a shared workspace.

Least privilege means giving each person and AI process the minimum access needed for the work, reducing what an error or compromised account could expose. Google’s access-level table helps when choosing a Shared Drive role; permissions differ between web editing and Drive for desktop. Restricted material needs enforced access controls, such as a limited-access folder or separate drive. A warning in AI instructions does not prevent access.

When the AI can change records or send messages

For a new workflow, starting with read access and drafts gives the team a chance to catch errors before they affect live records. One staged approach is:

1. Let the AI read approved project context.
2. Ask it to propose the task, update, message, or record.
3. Have a person check the source, recipient, and effect.
4. Add a narrowly defined write action after the workflow behaves well on representative examples.
5. Keep sending, publishing, deleting, moving money, and changing permissions behind explicit human approval unless the organization has reviewed and accepted a different control.

Specific standing permissions are easier to test. “The AI can update the CRM” leaves the records and actions undefined. For example: “The AI can read the funder’s name, application deadline, and last-contact date, then draft a follow-up in a Google Doc for review. It cannot edit the CRM or send the message.”

Visible status helps people know what they can use

AI makes it easy to create plausible material faster than people can review it. File names, folders, or workflow status can distinguish a draft from something reviewed or already published.

Possible states include:

• working or draft
• needs review
• approved
• sent or published
• superseded or archived

An approved draft is not proof that it was sent. A chat saying “done” is not proof that a live system changed. Linking the approved artifact to the live record helps when the distinction affects the work.

An adaptable project maintenance checklist

For shared or ongoing projects, the following checklist can help with setup, upkeep, and handover. Use the parts that fit the work and your organization’s policies.

When starting

• Name an owner and backup owner.
• Choose the project boundary and sensitivity level.
• Add a project brief with its purpose, instructions, and links to the current source documents.
• Choose access levels that match each person’s work.
• Test one real, low-risk task and ask which sources the AI used.

While active

• Record confirmed decisions outside chat history.
• Remove outdated uploads instead of leaving competing copies.
• Review access when people join, leave, or change roles.
• Check the project brief after a major decision or workflow change.
• Ask the AI to cite the source it used when current facts affect the output.

When closing

• Move approved outputs and decisions to the shared folders or systems your team maintains.
• Record the final status, owner, and remaining obligations.
• Remove temporary uploads and unnecessary connector access.
• Transfer or re-home the project content and access before a person leaves.
• Archive the project according to the organization’s retention policy.
• Test that another staff member can still find the durable record without the original owner’s account.

Archiving alone may preserve existing sharing permissions. Access that is no longer appropriate needs to be revoked separately.