What 'autopilot' has to earn: five rungs of autonomy
Handing an agent your ad budget is a big ask. The way we have tried to earn it is to make autonomy a ladder you climb deliberately, rather than a switch labelled "AI".
There are five rungs, and the default is the bottom one.
The ladder
- Off. The agent never touches the platform. It still scans, plans, generates creative and produces recommendations — you apply them.
- Auto-pause. It may pause underperforming ads. This is the safest possible first grant: the worst case is spending less money.
- Auto-pause and boost. It may also raise daily budgets on winners.
- Auto-pause, boost and tune. It may also shift bid strategy toward your CPA or ROAS targets — on existing campaigns only, reversibly, without creating new spend.
- Full autopilot. It may also draft new campaigns from proven winners.
Each rung is cumulative, and the executors check a capability predicate rather than comparing mode strings, so adding a rung happens in exactly one place.
Note what the top rung does not do
Even at full autopilot, the agent does not launch a campaign on its own. It drafts one and puts it in front of you. That is a deliberate ceiling: campaign creation is where a mistake is least reversible and hardest to notice, and we have not earned that yet.
Why moves are bounded
The budget orchestrator scores every active ad set on trailing 14-day ROAS, derives a pacing multiplier from month-to-date spend against the monthly budget, and reallocates accordingly — clamped to ±50% in any one run.
The bound matters more than the algorithm. A large reallocation converges over several daily runs instead of happening in one lurch, which means a single bad ROAS read cannot gut an ad set. Fourteen days is likewise a compromise: long enough to survive a quiet weekend, short enough to notice a genuine decline.
Every move is written to an audit trail whether it was applied or merely proposed, and bid-strategy changes snapshot the previous setting so they can be reverted.
The loop that makes it better
Each applied change records baseline metrics before and after. Those snapshots are rolled up per rule into win rate and observed lift, and the thresholds recalibrate against that. The point is not that the agent is clever on day one — it is that it accumulates evidence about your account rather than about accounts in general.
What this costs you in convenience
A ladder is slower than a switch. You have to decide what to grant, and the lower rungs will leave money on the table that a more aggressive system would have captured. We think that is the right trade while the evidence is still accumulating — and the ladder means you can move up whenever you decide it has earned it.