Deferral Policy

The primary mechanism for initiating a deferral is a confidence score falling below a calibrated threshold. The policy must define distinct thresholds for different risk classes.

• High-Risk Decisions: Defer if confidence < 99.9%
• Medium-Risk Decisions: Defer if confidence < 95%
• Low-Risk Decisions: Log only, no deferral This prevents the model from making guesses on edge cases and ensures selective prediction is enforced by business logic.

A deferral is not a simple stop. The policy must contain a decision tree that routes the task to the correct human role based on the nature of the uncertainty.

• Ambiguity: Route to a subject-matter expert for clarification.
• Safety Violation: Route to a safety officer and trigger a kill switch state.
• Compliance Flag: Route to legal and compliance teams for a risk acceptance sign-off. This logic prevents alert fatigue by ensuring the right person sees the right issue.

The policy must specify the maximum time a system can wait for human input before executing a fallback protocol. This prevents indefinite system stalls.

• Time-to-Respond (TTR): The window a human has to act (e.g., 15 minutes for financial trades).
• Time-to-Resolve (TTRs): The maximum total time for the human to fix the issue.
• Default Action: If the SLA expires, the system must revert to a safe state, such as canceling the transaction or applying a conservative default.

Every deferral event must generate an immutable audit trail to satisfy AI Audit Trail Immutability requirements. The log must capture:

• The specific input data that caused the low confidence.
• The model's raw output and confidence score.
• The identity of the human operator who accepted or overrode the decision.
• A timestamp of the human action. This data is critical for Continuous Compliance Monitoring and post-incident analysis.

The policy must define how a human's decision feeds back into the system. This is not just a one-way stop.

• Override: The human's decision replaces the AI's output for that specific transaction.
• Labeling: The human's action serves as a new ground-truth label for future Continuous Model Learning.
• Exception Handling: If a specific input pattern causes repeated deferrals, the policy should trigger an automatic review of the model's training data or prompt architecture.

A deferral policy must explicitly manage the system's Level of Automation (LoA) during the handoff. This prevents mode confusion.

• State Transition: The system must visually and programmatically signal a shift from 'Autonomous' to 'Awaiting Human Arbitration'.
• Input Locking: While awaiting human input, the system should lock the specific decision process to prevent conflicting automated actions.
• Teleoperation Readiness: For embodied systems, the policy must ensure a seamless handoff to a teleoperation interface without kinematic discontinuities.

A core routing mechanism that automatically escalates a decision to a human review queue when the AI model's prediction confidence score falls below a predefined, domain-specific boundary. This is the most common technical implementation of a deferral policy.

• Mechanism: A numerical threshold (e.g., < 95%) triggers a handoff.
• Calibration: Thresholds must be tuned per use case to balance automation rate against risk of error.
• Example: A medical imaging AI defers a diagnosis to a radiologist if its confidence in identifying a malignant tumor is below 98%.

An AI model's built-in capability to abstain from making a prediction on a specific input, triggering a deferral to a human expert when the model is uncertain. This is also known as learning to reject or classification with a reject option.

• Abstention: The model outputs "I don't know" instead of a low-confidence guess.
• Coverage: The proportion of inputs the model does make predictions on.
• Risk-Coverage Trade-off: A model can achieve higher accuracy on a smaller subset of inputs by deferring the rest.

A structured, hierarchical procedure that defines how an AI-generated issue or anomaly is progressively routed to higher levels of human authority based on severity, risk, or time sensitivity. A deferral policy specifies the trigger; the escalation protocol specifies the routing.

• Tier 1: Front-line reviewer for routine low-confidence deferrals.
• Tier 2: Subject-matter expert for complex edge cases.
• Tier 3: Executive or legal authority for high-risk or compliance-violating outputs.

A dynamic control paradigm where the level of autonomy transferred between a human operator and an AI system can be continuously adjusted along a spectrum in real-time based on task complexity. This contrasts with a binary deferral policy.

• Spectrum: Ranges from full manual control to full autonomy.
• Adaptive: The system dynamically shifts the Level of Automation (LoA) based on context.
• Example: A self-driving car handles highway cruising autonomously but smoothly transfers steering control back to the driver in a complex construction zone.

A predetermined, safe operational mode that an AI system automatically reverts to when it encounters an unexpected state or loses confidence, often involving a handoff to a human operator. This is the safety net when a standard deferral policy cannot execute.

• Safe State: The system defaults to a non-harmful, stable condition (e.g., halting a robotic arm).
• Degraded Mode: The system continues operating with reduced functionality and heightened human oversight.
• Example: A trading algorithm that loses connectivity to its primary data feed immediately cancels all open orders and alerts a human trader.

A cognitive bias where a human operator over-relies on an AI system's recommendation, ignoring contradictory information or failing to seek disconfirming evidence, even when the system is wrong. Effective deferral policies must account for this human factor.

• Risk: A reviewer may rubber-stamp a deferred decision without proper scrutiny.
• Mitigation: Interfaces should present the AI's confidence score and key evidence against its recommendation.
• Training: Operators must be trained to verify, not just validate, AI outputs.