AI Autonomy Levels Framework | 4-Stage Decision Automation by SAND
Platform / 02
Built for accountable autonomy.
Every Sand deployment begins at a level appropriate to the operator’s risk tolerance, data maturity, and regulatory environment.
Score a decision class
How would Symmetri handle your decision?
Adjust the four scoring dimensions below to see how Symmetri would classify a given decision class within the framework. The recommended automation level updates in real time as you change the inputs.
Decision parameters
Consequence Severity: Minor (3)
Worst-case outcome if the system acts incorrectly. Trivial → Catastrophic.
Model Confidence: Low (3)
Validated accuracy of the consequence model for this decision class right now.
Reversibility: Moderate (3)
Whether the action can be undone within an operationally acceptable window.
Decision Frequency: Moderate (3)
How often this decision occurs across the operational lifecycle.
Recommended automation level:
Human Required
Symmetri provides situational awareness only. No automated action under any circumstances.
Why: Moderate severity allows up to Supervised Automation provided confidence is sufficient. Current model confidence is below the threshold to reach that ceiling, so the gate restricts this class to Human Required.
The four automation levels
Each decision class lives at one of four levels.
Full Automation
System acts without human review
Action is executed directly. Outcomes are logged and audited downstream. Reserved for decision classes where the worst-case outcome is recoverable and the system has demonstrated sustained accuracy.
Requirements before deploying
- Consequence Model accuracy ≥85% for this decision class
- Action reversible within the operational window
- Severity low, worst case is recoverable
- Alert fatigue monitoring active
Typical decisions
- Burst alert generation
- Network fault notification
- Report scheduling
Supervised Automation
System executes, supervisor watches
Symmetri executes the action while a human supervisor has real-time visibility and one-click override at all times. The override window is technically enforced, not policy alone.
Requirements before deploying
- Consequence Model accuracy ≥75% for this decision class
- Override window defined and technically enforced
- Supervisor dashboard shows action rationale in plain language
- Defined escalation path if supervisor is unavailable
Typical decisions
- Pressure zone adjustment
- Traffic rerouting
- SOP-triggered workflow execution
Human-in-the-Loop
System recommends, human approves
Symmetri generates a ranked recommendation with full reasoning. A human reviews and approves before any action. The recommendation expires if not actioned within the approval SLA.
Requirements before deploying
- Counterfactual reasoning shown, why this action, not alternatives
- Confidence bounds displayed with every recommendation
- Defined approval SLA with expiry
- Audit trail of human decision and rationale
Typical decisions
- Drug procurement
- Facility capacity reallocation
- NOC escalation
- Maintenance scheduling
Human Required
System informs, human decides
Symmetri provides situational awareness and consequence modelling as briefing material only. No automated action under any circumstances. Used for high-consequence, irreversible decisions.
Requirements before deploying
- System surfaces all Consequence Model outputs as briefing material
- No automated action under any circumstances
- Decision rationale recorded in audit trail
- Regular review: does this classification still apply?
Typical decisions
- Emergency isolation
- Budget reallocation
- CapEx decisions
- Irreversible high-consequence actions
The four scoring dimensions
Every decision class is scored against four dimensions.
Dimension 01, Primary ceiling-setter
Consequence Severity
Worst-case outcome if the system acts incorrectly on this decision class.
Trivial → Catastrophic
Catastrophic severity = Human Required regardless of model accuracy.
Dimension 02, Primary gate
Model Confidence
How accurately the Consequence Model predicts outcomes for this decision class right now.
Unvalidated → Validated ≥85%
No class advances beyond Human-in-the-Loop until confidence ≥75% is validated.
Dimension 03, Ceiling modifier
Reversibility
Whether the action can be undone within an operationally acceptable window.
Permanent → Minutes to reverse
Low-severity but irreversible decisions are treated as higher severity.
Dimension 04, Efficiency multiplier
Decision Frequency
How often this decision occurs across the operational lifecycle.
Rare (annual) → Continuous
High frequency strengthens the automation case but does not lower the safety bar.
Worked examples
How real decisions land within the framework.
Across the domains where Symmetri is deployed, decision classes vary widely in their consequence profile. The same platform handles a routine burst alert and a multi-million-dollar capital decision, but they sit at very different automation levels.
Decision Classes Overview
| Decision class | Severity | Reversibility | Frequency | Level |
|---|---|---|---|---|
| Burst event detection alert | Low | Quick to reverse | Frequent | Full Automation |
| Demand forecast report generation | Trivial | Minutes to reverse | Continuous | Full Automation |
| Pressure zone valve adjustment | Moderate | Moderate | Moderate | Supervised Automation |
| Capital maintenance prioritisation | High | Permanent | Occasional | Human-in-the-Loop |
| Emergency network isolation | Catastrophic | Hard to reverse | Rare | Human Required |
| DECISION CLASS | SEVERITY | REVERSIBILITY | FREQUENCY | LEVEL |
|---|---|---|---|---|
| Stock-out risk alert to supply chain | Low | Minutes to reverse | Frequent | FULL AUTOMATION |
| Facility capacity utilisation report | Trivial | Minutes to reverse | Continuous | FULL AUTOMATION |
| Patient pathway escalation flag | Moderate | Moderate | Moderate | SUPERVISED AUTOMATION |
| Procurement order for essential drugs | High | Hard to reverse | Moderate | HUMAN-IN-THE-LOOP |
| National programme budget reallocation | Catastrophic | Permanent | Rare | HUMAN REQUIRED |
| DECISION CLASS | SEVERITY | REVERSIBILITY | FREQUENCY | LEVEL |
|---|---|---|---|---|
| Network fault consequence prediction | Low | Minutes to reverse | Continuous | FULL AUTOMATION |
| Traffic rerouting in congestion event | Moderate | Quick to reverse | Frequent | SUPERVISED AUTOMATION |
| NOC escalation trigger | Moderate | Quick to reverse | Moderate | SUPERVISED AUTOMATION |
| Infrastructure decommission recommendation | High | Permanent | Rare | HUMAN-IN-THE-LOOP |
| CapEx investment prioritisation | Catastrophic | Permanent | Rare | HUMAN REQUIRED |