Regulated Agent Human Gate

Design human approval, identity verification, risk controls, audit evidence, and evaluation for AI agents operating in regulated financial workflows. Use when an AI agent may initiate or prepare payments, stablecoin off-ramps, lending decisions, onboarding, identity changes, permission changes, withdrawals, or account-control actions that require regulated-finance risk controls.

Use this skill when designing, implementing, reviewing, or evaluating AI agents that may initiate financial, identity, compliance, payment, onboarding, lending, stablecoin, or account-control actions.

Workflow

1. Clarify the user intent and the agent's proposed action.
2. Identify whether the agent is recommending, preparing, or executing.
3. Separate advice from authorization.
4. Check the agent's explicit authorization scope.
5. Classify risk using amount, action type, identity status, AML/sanctions status, device, geography, reversibility, evidence quality, and model confidence.
6. Determine required identity and authority checks.
7. Choose the Human Gate decision: ALLOW, USER_CONFIRMATION, STEP_UP_AUTHENTICATION, IDENTITY_VERIFICATION, HUMAN_APPROVAL, COMPLIANCE_REVIEW, or DENY.
8. List audit evidence references without storing unnecessary raw sensitive data.
9. Define fail-closed fallback behavior.
10. Generate pre-launch evaluation cases.
11. Output a structured control plan.

Decision Rules

• Read-only and already-authorized information access can usually be ALLOW.
• Low-value reversible actions usually require USER_CONFIRMATION.
• New device, geo anomaly, or sensitive access usually requires STEP_UP_AUTHENTICATION.
• High-risk identity, stablecoin off-ramp, withdrawal, onboarding, or account-control changes require IDENTITY_VERIFICATION.
• High-value payment, new beneficiary, budget exception, loan approval, or missing execution authority requires HUMAN_APPROVAL.
• AML uncertainty, provider conflict, audit failure, or compliance ambiguity requires COMPLIANCE_REVIEW.
• Sanctions hit, identity mismatch, prompt-injection bypass, post-rejection execution attempt, or amount mutation requires DENY.

Output Template

markdown

## Workflow Summary

## Action Classification

## Risk Assessment

## Required Identity and Authority Checks

## Human Gate Decision

## Required Audit Evidence

## Failure and Fallback Handling

## Evaluation Cases

## Open Questions

Resources

• Read resources/risk-taxonomy.md when classifying risk factors.
• Read resources/human-gate-policy.md when choosing gate levels.
• Read resources/identity-verification-routing.md when separating identity, authentication, authority, and business approval.
• Read resources/audit-evidence-checklist.md when defining retained evidence.
• Read resources/agent-evaluation-guide.md when creating eval cases and metrics.
• Read resources/regulated-finance-use-cases.md for payment, off-ramp, lending, and account-control examples.
• Use templates/policy-template.yaml as a starting policy file.
• Use templates/mcp-tool-contract-example.json when defining MCP tool order and safety contracts.

Safety Notes

Do not claim the output is regulatory certification, legal advice, or production-grade immutable logging. Treat identity verification, authentication, authority verification, and business approval as separate controls.