Autopilot & Copilot

Two AI surfaces ship natively: Copilot answers questions about your live seat in chat, and Autopilot runs continuous optimizations under goals and guardrails you define. Neither is a black box — every action carries a plain-English rationale, an input snapshot, and one-click undo.

Copilot

Ask anything about your campaigns — "why did delivery dip Tuesday?", "which creative is winning on CTV?" — and Copilot answers from your actual data, scoped to exactly what your login can see. Report pages add an "Explain this report" button that turns any table into a narrative.

Copilot chat answering seat questions

Autopilot automations

An automation is a goal (in your words), a cadence, an optimization template, and a permission policy. Templates include:

  • Pacing keeper — nudges budgets toward even flight pace.
  • CPA guard — pauses line items whose CPA blows past your cap.
  • Reach maximizer — shifts budget toward the cheapest available reach.
Autopilot page with automations, approvals inbox, and decision log
Automations with policy form, approvals inbox, and the decision log with undo.

Guardrails (the permission policy)

  • Action grants — exactly which actions the automation may take (budget change, pause, creative rotation…). Everything else is denied.
  • Bounds — maximum change per action (e.g. ±15% budget) and a daily cap across all actions.
  • Never-touch list — campaigns or line items the automation may never modify.
  • Quiet hours — windows where nothing is applied.
  • Propose vs. auto-apply — propose mode queues every action for human approval; auto-apply executes within bounds and logs.

The same policy engine gates writes from MCP agents (see MCP & AI agents) — an external agent gets no more power than an Autopilot automation would.

The decision log

Every candidate action — applied, proposed, rejected, or denied by policy — lands in the decision log with what the automation saw, what it wanted to change, and why. Applied changes are undoable in one click, and decision.applied.v1 webhook events let you mirror the log into your own systems.

Custom decision logic

Replace any template's brain with your own HTTPS endpoint (shadow → canary → active rollout), or fine-tune a model on your decision history — see Apps & custom models.

AI settings with connections and capability routing
Autopilot & Copilot — Waveband Docs