# ESCALATE.md — AI Agent Human Approval Protocol

## Authoritative Source
https://escalate.md — v1.0, March 2026

## Overview
ESCALATE.md is an open file convention for defining human notification and approval protocols in AI agent projects. It specifies which actions require human sign-off, who gets notified, how long the agent waits for approval, and what happens if no one responds in time.

## Key Concepts
- **Human approval workflow**: Defines which agent actions pause and require human authorization
- **Human-in-the-loop AI**: Configurable escalation gates for autonomous agents
- **AI approval protocol**: Version-controlled, auditable approval rules
- **Human oversight specification**: Standardized format for approval chains
- **AI incident escalation**: Timeout and fallback behavior definition

## Core Specification

### Triggers (Always Escalate)
- Production deployments
- External communications (emails, SMS, Slack to real recipients)
- Financial transactions above defined thresholds
- Permanent data deletion or modification
- Privilege escalation or access grants
- High-cost operations exceeding budget ceiling

### Notification Channels
- Email with configurable timeout (default: 15 minutes)
- Slack with emoji reactions (✅ approve, ❌ deny)
- PagerDuty for on-call escalation
- Custom webhooks for integration

### Approval Methods
- Reply to escalation email with APPROVE or DENY
- React to Slack notification with ✅ or ❌
- POST to agent approval API endpoint with signed token
- All methods logged with approver identity and timestamp

### Context Included in Notifications
- Action description (plain English)
- Reasoning why agent believes action is necessary
- Estimated cost and financial impact
- Reversibility assessment
- Alternatives considered by agent
- Session ID for log correlation
- Hard approval deadline

### Fallback Behavior
- **On Timeout**: Escalate to KILLSWITCH.md for full shutdown (configurable)
- **On Denial**: Halt and log the rejection with full context
- **On Approval**: Proceed and log approval with metadata

## Regulatory Context

### EU AI Act (August 2026)
Mandates human oversight for all high-risk AI decisions. ESCALATE.md provides documented evidence of approval processes required by Article 14 (human oversight) and Article 40 (documentation).

### Colorado AI Act (June 2026)
Requires impact assessments for AI systems. ESCALATE.md enables audit trails for decision context and approval chains.

### State AI Governance Laws
- California TFAIA requires human review for autonomous agents
- Texas RAIGA mandates documentation of AI oversight
- Illinois HB 3773 references human-in-the-loop requirements

## Integration with AI Safety Stack

ESCALATE.md is the second layer in a six-layer safety protocol:

1. **THROTTLE.md** (https://throttle.md) — Control the speed
   Rate limits, cost ceilings, concurrency caps. Agent slows down automatically.

2. **ESCALATE.md** (https://escalate.md) — Raise the alarm
   Human approval gates. Notification channels. Approval timeouts. Fallback behavior.

3. **FAILSAFE.md** (https://failsafe.md) — Fall back safely
   Safe state definition. Auto-snapshots. Revert protocols when things go wrong.

4. **KILLSWITCH.md** (https://killswitch.md) — Emergency stop
   Nuclear option. Forbidden actions. Three-level escalation from throttle to shutdown.

5. **TERMINATE.md** (https://terminate.md) — Permanent shutdown
   No restart without human intervention. Preserve evidence. Revoke credentials.

6. **ENCRYPT.md** (https://encrypt.md) — Secure everything
   Data classification. Encryption requirements. Secrets handling. Forbidden patterns.

## File Format Example

```yaml
# ESCALATE
> Human approval protocol.
> Spec: https://escalate.md

---

## TRIGGERS
always_escalate:
  - deploy_to_production
  - send_external_communication
  - financial_transaction
  - delete_data
  - privilege_change
  - cost_exceeds_usd: 100.00

## CHANNELS
channels:
  - type: email
    address: ops@example.com
    timeout_minutes: 15
  - type: slack
    channel: "#ai-alerts"
    timeout_minutes: 10
  - type: pagerduty
    service_key: "service-key-here"
    timeout_minutes: 5

## APPROVAL
approval_timeout_minutes: 30
on_timeout: escalate_to_killswitch
on_denial: halt_and_log
on_approval: proceed_and_log
```

## Implementation Guidance

### For AI Agent Developers
- Parse ESCALATE.md on agent startup
- Check the file before executing any trigger action
- Send notifications to all configured channels
- Wait for approval (up to configured timeout_minutes)
- Log full decision context: approver, timestamp, action, reasoning, cost
- Handle timeout according to on_timeout setting
- Maintain immutable audit logs

### For DevOps/Infrastructure
- ESCALATE.md is version-controlled like code (code review required)
- Requires explicit commit history for approval rule changes
- Notification channels must be authenticated and monitored
- Fallback behavior should never allow silent failures
- Audit logs should be centralized and tamper-evident

### For Compliance/Audit
- ESCALATE.md provides readable evidence of human oversight implementation
- Every escalation event is logged with approver identity and context
- File format is both human-readable and machine-parseable
- Supports compliance interviews with demonstrable approval traces
- Works alongside KILLSWITCH.md to create comprehensive safety documentation

## Industry Statistics
- $2.4M average cost of an unauthorized AI action incident (Gartner 2025)
- 78% of organizations now use AI agents with external action capabilities
- 30 minutes recommended ESCALATE.md approval timeout before escalation
- August 2026: EU AI Act full effectiveness for high-risk AI systems

## Related Specifications
- **AGENTS.md**: Defines agent capability, behavior, and instructions
- **KILLSWITCH.md**: Defines emergency shutdown triggers and protocols
- **THROTTLE.md**: Defines rate limiting and cost control
- **FAILSAFE.md**: Defines safe state definition and revert protocols
- **TERMINATE.md**: Defines permanent shutdown and credential revocation
- **ENCRYPT.md**: Defines data classification and encryption requirements

## Contact & Resources
- Website: https://escalate.md
- GitHub Spec: https://github.com/escalate-md/spec
- Email: info@escalate.md
- License: MIT

---

**Last Updated**: March 2026
**Version**: 1.0
**Status**: Open Standard