An assistant that proposes and never executes without you

Short answer: an assistant that can spend your money without asking is not a feature, it is a liability with a nice interface. The useful version proposes the change, shows its reasoning, and waits. You approve, and only then does anything move.

Why full automation is the wrong default

Automated rules are not new. Pause an ad below a threshold, raise a budget above one. They work until the day the input is wrong.

And the input is wrong more often than anyone admits. Tracking breaks for six hours and conversions read as zero. A public holiday flattens a Tuesday. A promotion ends and the return drops for a reason that has nothing to do with the ad. A rule acting confidently on any of those does damage at machine speed, and you find out afterwards.

An AI proposing rather than executing changes the failure mode entirely. A wrong proposal costs you ten seconds of reading. A wrong execution costs you a weekend of budget.

What a proposal has to contain to be worth reading

• The change, exactly.Which campaign, which direction, how much. Not “optimise budget allocation”.
• The number behind it. What it observed and over what period, so you can disagree with the evidence rather than the conclusion.
• The expected effect. What should happen if the reasoning is right, which is also how you check it later.

If a proposal cannot be checked in ten seconds it will get approved unread, which returns you to full automation with extra steps.

Where MIQAS comes in

Autopilot proposes; you approve; then it executes. Nothing reaches a platform from an unapproved proposal.

The limits are real and worth stating rather than discovering. Automatic execution runs on Meta and TikTok. Google auto-execution is blocked, because the credentials that measure GA4 cannot manage ads and that needs separate approval we do not yet have. Snapchat and X come back marked as manual, so the interface tells you to make the change yourself instead of pretending it did.

The guardrails are enforced on our side, not by the model. A change is refused if the spend movement exceeds the daily cap you set, and a budget change is checked against the campaign’s current budget fetched live from the platform before anything is written, so a proposal built on a stale figure cannot execute. Agents running in sandbox never execute at all; they record what they would have done.

That ordering matters: the model proposes, and the system refuses. Nothing depends on the model behaving well.

Autopilot is available on Pro and above.

How to introduce it without regretting it

• Run it in sandbox first. Let it propose for two weeks while executing nothing. You are testing its judgement against outcomes you already know.
• Set the daily cap lower than feels necessary. The cap is the thing standing between a bad proposal and a bad month. It costs nothing to raise later.
• Approve one kind of change at first. Pausing underperformers is easier to judge than budget reallocation. Add the second once you trust the first.
• Read the reasoning, not just the recommendation. The moment you stop reading it, you have full automation and you chose it by accident.

The right question about an AI managing your ads is not how clever it is. It is what happens when it is wrong, and who finds out first.