Escalation Model

ASBOS uses three action tiers. Not all decisions are fully autonomous.

Tier 1: Low-risk and reversible

Examples: drafting a schedule, sending an approved reminder, routine non-sensitive operational updates.

Tier 2: Autonomous with notification

Examples: confirming a replacement within a pre-approved pay range, actions that may execute automatically but notify a supervisor.

Tier 3: Human approval required

Examples:

• Unbudgeted premiums
• Adverse personnel actions
• Payroll changes
• High-severity incident communications
• Credential exceptions
• Unfillable coverage gaps
• Suspected fraud
• Client complaints and legal or safety-sensitive incidents

Voice, privacy, and safety

Voice deployment requires jurisdiction-specific review of call authorization, recording consent, AI disclosure, data retention, and opt-out handling.

• The Federal Trade Commission telemarketing and prerecorded-call rules may apply depending on call type and jurisdiction.
• Illinois BIPA treats voiceprints as biometric identifiers. ASBOS should not create biometric templates unless the use case is necessary and legally approved.
• The safest default is ordinary audio recording only when approved — not speaker recognition or voice authentication without legal review.

Research metrics

A publishable pilot should track:

• Human escalation rate
• False-positive exception rate
• Percentage of events resolved at each tier
• Error rate of autonomous decisions compared across systems