ROASfinity Get early access

MCP permissions for agencies: a practical model

The question is not whether to let assistants near client accounts. It is which hands they get and who signs off.

Roles, not tools

Reader: can look at data and explain it. Default for every new connection.
Drafter: can write proposals that a person reviews. Nothing is sent.
Operator: can apply an approved change. Needs a named approver for each change.

Scopes per client

Give each connection access to the clients it needs and no more.
Keep one client's data out of another client's context.

Guard rails

Start every new ad paused.
Cap budget moves per change and per day.
Require a read-back after every applied change and flag any difference.

Logging

Log the assistant that proposed, the person who approved, the exact change, the time and the read-back.
Make the log exportable, and keep it for as long as the client relationship lasts.

Review

Every quarter, list the connections, who owns each one and what they can do. Remove what nobody can explain.
After any incident, tighten the scope first and loosen it later, with a reason.

The AI assistant safety checklist is the short form of this model.

7 min read. Last updated 2026-10-07. Examples use made-up numbers.

This guide in numbers

7minutes to read

3terms explained

5sections

6related guides

Use it as a checklist

Tick items as you go. Your ticks stay in this browser only.

Reader: can look at data and explain it. Default for every new connection.
Drafter: can write proposals that a person reviews. Nothing is sent.
Operator: can apply an approved change. Needs a named approver for each change.

Terms in this guide

MCP Approval workflow Audit trail

Keep reading

AI and automation

A safety checklist before an AI assistant touches ad accounts

An assistant that can read your accounts is a convenience. One that can write to them is a decision. This checklist separates the two.

6 min read AI and automation

AI assistants and ad accounts: read, propose, apply

Connecting an assistant to an ad account is easy. Deciding what it may do without asking is the part that needs a rule.

7 min read Operations

A client approval checklist for ad launches

Client approval goes wrong in the same ways: the wrong version is approved, the answer lives in a chat, and nobody can say later what was agreed.

6 min read Operations

A client offboarding checklist for agencies

A clean exit protects both sides. It also leaves the door open for the client to return.

5 min read Operations

A process for requesting access to client ad accounts

Every login an agency holds is a risk and a responsibility. A written request, a named approver and a planned removal keep the list short and explainable.