Inputs
The engine consumes authorized campaign, search, analytics, and product context that is available to the requesting workspace. Missing, stale, or failed sources reduce the evidence available to a recommendation and must not be silently replaced.
- Synced provider metrics and dates
- Campaign and account context
- Detected anomalies or rule triggers
- User feedback, status, and prior action history where available
Rules and AI involvement
Deterministic checks can identify conditions such as missing data, changes in spend, or a threshold breach. AI services can summarize the evidence, rank candidate next actions, and explain why a suggestion matters. Generated language is not treated as ground truth; source metrics and execution validation remain separate.
Human approval and execution
A recommendation can be approved, rejected, or left for review. Management-capable actions are executed only through authenticated, owned, action-specific endpoints after the user initiates the action. Connector validation, idempotency, error handling, and audit state protect the provider boundary.
Worked demo-data example
A fictional campaign shows rising cost per lead and declining lead volume. The engine can present “review budget and creative fatigue” with the relevant demo metrics. The user inspects tracking quality and business context before approving a supported action or creating an experiment. No improvement is promised.