Skip to main content

​
Overview

The Decisioning Autopilot watches your experiments and models in the background and proposes concrete configuration changes. Every proposal lands in the Recommendation Inbox (AI > Insights) as an AiRecommendation; what happens next is controlled by a per-tenant autonomy mode. The decision path itself stays fully deterministic — the Autopilot only tunes configuration, through the same review and audit machinery a human would use.

​
What it proposes

The Autopilot sweep runs every 6 hours (GET /api/v1/cron/ai-autopilot, CRON_SECRET-gated, registered in the in-process maintenance scheduler). Per tenant it produces two kinds of proposals:
ProposalSourceTrigger
Promote challenger (type: experiment)evaluateAutoPromote() over active experiments with autoPromote: trueChallenger beats the champion with statistical significance: AUC improvement ≥ promoteThreshold, Bonferroni-corrected p-value, ≥ 5% relative lift, ≥ 100 samples per variant, and promoteAfterDays elapsed
Retrain model (type: model)detectModelDrift() over active modelsDrift severity significant (AUC dropped > 0.1 from peak)
Proposals are deduplicated: a re-run never creates a second open recommendation with the same type and title.

​
Autonomy modes

Configured in Settings > AI Configuration > AI Autonomy (stored on tenant_settings.aiAutopilot, editable via PUT /api/v1/tenant-settings with {"aiAutopilot": {"mode": "..."}}):
ModeBehavior
suggest (default)Proposals wait in the Recommendation Inbox for an admin to Apply or Dismiss.
auto_gatedAdditionally opens an Approval Request per proposal (entityType: ai_recommendation, requester autopilot). Four-eyes reviewers see it in Settings > Approvals.
autoApplies each proposal immediately through the shared apply executor. Every application is audit-logged (ai_recommendation / autopilot: true).

​
Applying a proposal

Whether applied by hand or by the Autopilot, all applications flow through the same executor:
  • experiment — flips championModelId on the (tenant-scoped) experiment and sets it active — the same operation as the playbook_promote_challenger_if_winning playbook.
  • model with action: retrain — runs the shared retrain-and-persist path (same as POST /api/v1/algorithm-models/[id]/train): version snapshot, metrics history, model update.
  • model with action: pause — sets the model status to paused.
  • weights — updates a ranking profile’s weights and bumps its version.
A failed application returns an error and the recommendation stays open — a recommendation is never marked applied unless the change actually happened.

​
Relationship to scheduled retrains

The existing scheduled-retrains cron (auto-retrain on autoLearn models) is unchanged. The Autopilot adds drift-triggered retrain proposals and the experiment promotion loop on top — the promotion evaluator (evaluateAutoPromote) previously existed as configuration only, with nothing running it.