Trust & Autonomy
AtlasAI uses a progressive autonomy model that lets organizations control how much authority the AI has over operational actions. The system is designed to build trust incrementally — starting with AI as an advisor and gradually allowing more autonomous action as the team gains confidence.
Autonomy Levels
| Level | Name | AI Behavior | Human Role |
|---|---|---|---|
| L0 | Inform | Reports findings and observations | Human investigates and acts on everything |
| L1 | Suggest | Suggests specific actions with reasoning | Human reviews and approves each action |
| L2 | Act & Report | Executes low-risk actions automatically, suggests medium/high-risk | Human approves medium/high-risk actions |
| L3 | Act & Alert | Executes low and medium-risk actions, alerts on high-risk | Human handles high-risk actions only |
| L4 | Full Auto | Handles incidents end-to-end, including high-risk actions | Human reviews after resolution |
| L5 | Closed Loop | Resolves and closes incidents without human intervention | Human reviews periodically via reports |
How Trust Levels Apply
Trust levels can be configured at multiple granularities:
- Global default — The baseline trust level for all services and action types
- Per service — Override the global level for specific services (e.g., L3 for staging, L1 for production databases)
- Per action type — Set different levels for different operations (e.g., L2 for restarts, L0 for data deletion)
- Per time window — Increase autonomy during business hours when the team is available, reduce during off-hours
Risk Classification
Every action the AI proposes is classified by risk:
| Risk Level | Examples | Default Behavior |
|---|---|---|
| Low | Read-only diagnostics, metric queries, log searches | Auto-execute at L2+ |
| Medium | Service restarts, scaling operations, cache flushes | Auto-execute at L3+ |
| High | Database failovers, production deployments, data migrations | Auto-execute at L4+ |
| Critical | Data deletion, credential rotation, network changes | Always requires approval (even at L5) |
Configuring Trust Levels
- Navigate to Settings → AI → Autonomy
- Set the Global Default level (recommended: start at L1)
- Add Service Overrides for services that need different trust levels
- Add Action Overrides for action types that need specific handling
- Review the Trust Dashboard to see current autonomy coverage
Building Trust Over Time
AtlasAI tracks AI accuracy metrics that help you decide when to increase autonomy:
- RCA accuracy — Percentage of RCA results marked correct by operators
- Runbook success rate — Percentage of AI-generated runbooks that resolved the incident
- False positive rate — How often the AI suggested unnecessary actions
- Time saved — Estimated hours saved by AI-automated actions vs. manual resolution
These metrics are available under Settings → AI → Trust Dashboard. When accuracy metrics are consistently above your comfort threshold, consider increasing the autonomy level by one step.
Safety Guardrails
Regardless of trust level, these guardrails always apply:
- Rate limiting — Maximum number of automated actions per hour (configurable)
- Blast radius limits — AI cannot execute actions affecting more than N services simultaneously
- Kill switch — Instantly disable all AI automation from the Command Center
- Escalation — If AI confidence is below threshold, it always escalates to a human
- Critical action lock — Destructive operations always require human approval