Autopilot is not a global switch. Autonomy in the Marketing Brain is granted per lever — a specific action type in a specific domain — and it's earned from that lever's own measured record on your store, not from a demo or a default. A lever that's been right about budget nudges says nothing about whether it should touch prices.
Fast down, slow up
- Promotion is slow: a lever needs at least 10 decided actions with a win rate of 75% or better before it can be promoted past operator review.
- Demotion is fast: a poor win rate or a loss streak demotes the lever back to operator review immediately. It re-earns autonomy the same way it did the first time.
- The asymmetry is deliberate. Losing autonomy should be cheap and instant; gaining it should require a record.
The verdict engine
The record comes from the verdict engine, which scores every applied action win, loss, or neutral — after the full measurement window, not on day-two noise. Wins are judged against your store's break-even ROAS, with a baseline-drift adjustment so a rising or falling market doesn't get credited to (or blamed on) the action. Auto-rollback of confirmed losers exists, but it's double-gated: you opt in, and each rollback still respects its own gate.
What never runs alone
Irreversible actions are never auto-approved — at any win rate, in any mode. If it can't be cleanly undone, it waits for a human, permanently. And confidence calibration is one-way-down: a realized win rate can add review requirements to a lever, but nothing the system measures can ever remove a review requirement on its own.
The budget envelope
- Each domain has a soft daily spend cap. A budget increase that would breach the envelope is blocked, autonomous or not.
- Decreases and pauses are always allowed — spending less never needs permission.
- Individual budget moves are clamped to ±20%, so even an earned, in-envelope change can't lurch.
The ladder means you're not trusting 'the AI'. You're trusting a specific lever's specific record, which you can read.