Governance, Security & Human Control

Seven principles of enterprise AI governance

The mission is grounded in seven operating principles. Each one specifies who holds authority, what evidence is required, and how failure is handled.

1. Authority Must Be Explicit

The organization defines what the agent may read, what it may draft, what it may recommend, what it may execute, which thresholds apply, which approvals are required, and which actions are prohibited.

2. Human Control Must Be Meaningful

A human review is meaningful only when the reviewer has authority, sufficient context, enough time, a clear decision, visible risk, an alternative, the ability to reject, and the ability to stop execution.

3. Least Privilege Is the Default

Workers receive only the access needed for their role and current workflow.

4. Risk Determines Control

Higher-risk work requires stronger authentication, evidence, review, segregation of duties, monitoring, rollback, audit, and incident response.

5. Evidence Must Be Durable

The system preserves input, sources, worker, model, tool, approval, action, result, correction, incident, and outcome.

6. Failure Must Be Visible and Recoverable

The system stops, escalates, or degrades safely.

7. Governance Must Improve Through Operations

Governance findings enter the Engineering Loop so controls strengthen as the workforce operates.

Governance architecture

Governance connects strategic oversight to operational controls. The cascade begins with the Board or Executive Oversight, which delegates to the AI Governance Council. The Council is supported by six functions — Business Owner, Security and Privacy, Legal and Compliance, Technology and Architecture, Risk and Internal Audit, and Human Resources or Workforce Governance — and is implemented through the Agent Product Owner.

The Agent Product Owner owns the operational controls that turn policy into enforced behavior: the Role and Authority Registry, Risk Tiering, Policy Engine, Approval Model, Monitoring and Audit, Incident Response, and the Engineering Loop. Each control is the runtime artifact that binds a principle to the system in motion.

The supporting Council functions map to the nine governance roles: Board or Executive Oversight, AI Governance Council, Business Owner, Agent Product Owner, Security and Privacy Owner, Legal and Compliance Owner, Technical Owner, Model and Evaluation Owner, and Human Reviewer. The council cross-checks use-case approval, risk-tier standards, policy, exception review, incident oversight, periodic review, deprecation, and portfolio reporting.

Authority levels

Authority is represented as a ladder. Each step widens the surface the agent can act on and narrows the human attention required. The ladder ends at a hard prohibition boundary.

1. Level 0 — Observe. The agent may read approved information, summarize, classify, detect, and report.
2. Level 1 — Draft. The agent may prepare communication, work items, decision packages, configuration, and code. A human executes or approves.
3. Level 2 — Recommend. The agent may compare options, recommend action, identify risk, and propose prioritization. A human decides.
4. Level 3 — Execute Low-Risk Work. The agent may perform approved, reversible, routine actions — for example, creating an internal task, scheduling a reminder, updating a non-sensitive status, or sending an approved transactional notification.
5. Level 4 — Execute With Approval. The agent may execute after authorized approval — for example, issuing a refund, deploying a production release, modifying access, sending external legal communication, or approving a contract workflow step.
6. Level 5 — Prohibited. The agent may not perform the action. Examples include autonomous termination, autonomous legal representation, autonomous safety waiver, autonomous high-impact eligibility decisions, autonomous criminal accusation, and uncontrolled destructive action.

Risk tiers

Risk is assessed across fourteen dimensions: customer or employee impact, financial impact, safety, legal or regulatory impact, privacy, discrimination, reversibility, scale, public visibility, system criticality, data sensitivity, fraud or abuse, model uncertainty, and dependency risk. The assessed tier then drives the controls.

Tier 1 — Low

Public information, internal summaries, reminders, non-sensitive status updates, and drafts. Controls: standard monitoring, routine testing, and periodic review.

Tier 2 — Moderate

Account-specific support, operational routing, standard returns, routine record updates, and internal recommendations. Controls: authentication, bounded tools, human fallback, quality sampling, and audit.

Tier 3 — High

Financial actions, sensitive employee records, legal matters, production deployments, access changes, and high-value customer actions. Controls: strong authentication, human approval, segregation of duties, evidence package, enhanced monitoring, and rollback.

Tier 4 — Critical or Prohibited

Safety waivers, autonomous employment termination, autonomous criminal or legal determinations, uncontrolled destructive action, and high-impact eligibility without authorized review. Controls: prohibit or require a specialized human process.

Human control patterns

Different work demands different relationships between agent and human. Five patterns cover the surface from routine bounded work to high-stakes review.

• Human in the Loop. A human approves before execution. Use for high-risk writes, legal interpretation, sensitive communication, and material financial action.
• Human on the Loop. The agent executes routine bounded work while humans monitor, sample, and intervene. Use for high-volume low-risk tasks, standard notifications, and routine internal work.
• Human Over the Loop. Humans define policy, thresholds, and system boundaries, and review outcomes. Use for mature, stable, low-risk operations.
• Human Takes Over. The agent transfers the active workflow to a person. Use for sensitive matters, ambiguity, emotional escalation, tool failure, user request, or policy exception.
• Two-Person Control. Two authorized people approve. Use for large financial transfers, highly sensitive access changes, production destructive action, material public communication, and critical legal action.

Role charter

Every Digital Worker has a role charter. The charter binds the abstract principles to a specific role and is the artifact the Agent Product Owner reviews.

The charter carries sixteen required fields: role name, business purpose, owner, users served, allowed inputs, allowed data, allowed tools, allowed actions, prohibited actions, approval requirements, escalation conditions, success metrics, quality thresholds, stop conditions, retention, review cadence, and version.

Example for a Refund Operations Worker at Authority Level 2:

role: Refund Operations Worker
authority_level: 2
purpose: Prepare refund recommendations for valid commerce orders
allowed:
  - retrieve_order
  - retrieve_return
  - calculate_policy_refund
  - create_refund_review_package
prohibited:
  - execute_refund
  - modify_fraud_status
  - override_return_policy
human_owner: Customer Operations Manager
review_required:
  - refunds_above_100_usd
  - fraud_indicator
  - disputed_delivery

The charter is the bridge between the seven principles and the operational controls. Identities, approvals, evidence, and rollback all derive from the charter's fields.

Stub-content markers have been replaced as the source chapter is now in place.