An approval queue, not an autopilot
Every proposed change waits for a human. There is no setting that bypasses this, and we will not add one without asking you first.
AI Actions
It does the work. You keep the decision.
“Can it just fix this for me?”
Finding a leak is only half the job. The other half is the twenty minutes of tab-switching, the ad account you opened in the wrong profile, and the budget you typed one zero wrong. AI Actions closes that gap: the model proposes a specific change with its evidence attached, you approve it, and it executes against the platform's API in about a second. Nothing reaches your ad account without that approval, and every action is logged and reversible.
Open AI Actions in the demo →Sample data · no signup
What you actually see
Executed over typed MCP tools
Meta and Google are exposed to the agent as permissioned, typed tools — pause a campaign, set a budget, adjust a bid cap. It cannot invent an action that does not exist.
An immutable audit log
Who proposed it, what evidence it cited, who approved it, when it shipped and what the platform answered. Written down permanently.
One-click revert
Every executed change knows its previous state, so undoing it is a click rather than an archaeology exercise.
How it gets there
- 1
The model proposes
Running on your resolved profit data rather than on a screenshot of it. The proposal carries the numbers it was based on.
- 2
You approve, or you do not
Declining is a first-class outcome and is logged the same way. A tool that makes refusal awkward is not asking a question.
- 3
MCP executes and confirms
The change goes to the platform's own API, and the response is stored with the action. If the platform rejects it, you see why.
Where it stops
What AI Actions will not tell you
- It acts only within the tools it has been given. It cannot create campaigns, change creative or touch billing.
- Approval is per change, not a blanket permission. A large batch is a large number of decisions, deliberately.
- Platform APIs have their own rules and rate limits. Some changes take effect on the platform's schedule, not ours.
▸Can it change my campaigns without asking?
No. That is a contractual commitment in our terms, not a preference in a settings page. Every write is a proposal until a human approves that specific change.
▸What is MCP?
The Model Context Protocol — an open standard for giving an AI typed, permissioned access to real systems. It matters here because the agent gets a defined set of actions rather than a browser and a password.
▸What if it makes a mistake?
Every action stores the state it replaced, so reverting is one click. And because approval is required, a mistake has to get past you first.
The module next to it
All six modules run on sample data in the open demo — no signup. Open Kepra →