PCC Learning Center · AI Governance
Decide What AI Can Do Alone—and What a Person Must Approve
Use consequence, reversibility, data sensitivity and authority to decide where AI needs routine review, explicit approval or a firm boundary against automation.
Quick Rule
Require approval before AI takes an action that affects people, money, legal obligations, security, sensitive data or a difficult-to-reverse business process.
A Person Checks the Output
Appropriate for lower-risk work such as draft internal communications, meeting summaries or suggested spreadsheet formulas.
The Workflow Pauses Before Action
An authorized person approves or rejects the proposed action before it sends, changes, purchases, grants access or deletes anything.
Microsoft describes human-in-the-loop approval as pausing an agent’s tool execution until a person approves the proposed action. Microsoft: Human-in-the-Loop with AG-UI →
Four Questions Determine the Level of Oversight
Ask these questions before allowing an AI action to proceed automatically.
1. What Happens If It Is Wrong?
Minor rework is different from financial, security, contractual or reputational damage. More serious consequences require stronger approval.
2. Can It Be Reversed?
A draft can be edited. Disclosing confidential information, sending a large external message or permanently deleting records may be impossible to undo.
3. What Data or Authority Is Involved?
Read-only access to public information creates less risk than permission to access sensitive records, change systems, send, purchase or delete.
4. Who Is Affected?
Actions materially affecting clients, employees, applicants, vendors or the public deserve stronger scrutiny—especially when outcomes are difficult to contest.
Human-Approval Decision Matrix
Use this as a starting point, then adjust thresholds for your industry, contracts, regulatory obligations and risk tolerance.
| Oversight level | Use when | Examples | Required control |
|---|---|---|---|
| Routine review | Consequences are minor, reversible and do not involve a sensitive action. | Internal drafts, formatting, meeting topics | User checks the output; repetitive work may use periodic quality sampling. |
| Explicit approval | The action affects people, money, security, sensitive data, commitments or business records. | Client messages, credits, record changes, access grants | Workflow pauses and an authorized person sees context and impact before deciding. |
| Two-person or specialist approval | Potential impact is high, regulated or concentrated. | Large payments, privileged access, bulk data changes, legal commitments | Require a specialist or two reviewers and preserve the supporting evidence. |
| Do not automate | Meaningful review is impossible or consequences exceed the business’s risk tolerance. | Unsupported safety decisions, deceptive impersonation, prohibited actions | Remove the required tool or permission and route the task to a qualified person. |
Actions That Usually Require Approval
External Communications
Messages that create promises, instructions, admissions, pricing decisions or significant client expectations.
Money and Commitments
Payments, refunds, pricing exceptions, bank-detail changes and commitments above defined thresholds.
Security and Access
Account creation, access grants, role changes, credential operations and security-policy exceptions.
Important Records
Bulk changes, deletions, status updates or changes that trigger billing, notifications or downstream workflows.
Decisions Affecting People
Consequential employment, eligibility, discipline, performance or access-to-service decisions.
Sensitive Information
Sending, publishing or sharing confidential, personal, regulated or security-sensitive information outside its intended boundary.
What a Meaningful Approval Request Must Show
A generic “Approve” button is not enough. The reviewer needs enough source context to make an independent decision.
- What the AI proposes to do
- Who, what system or which audience will be affected
- The relevant source information
- Important assumptions or uncertainty
- What data will be disclosed or changed
- Whether the action can be reversed
- The deadline or business reason
- The requesting user and acting agent
- Options to approve, reject or return for correction
Build Approval Into the Workflow
Telling employees to “double-check AI” is not a reliable control for high-impact actions.
- The AI prepares the proposed action and supporting context.
- The workflow pauses before execution.
- The appropriate reviewer receives the request.
- The reviewer approves, rejects or requests a correction.
- The system records the decision and reviewer identity.
- Only an approved action proceeds.
- Exceptions and failures are monitored.
Document the Rule Before Deployment
For every AI use case, record these decisions in the AI Tool Inventory or the agent’s operating documentation.
- Actions AI may take automatically
- Actions requiring routine review
- Actions requiring explicit pre-approval
- Actions requiring specialist or two-person approval
- Actions AI is prohibited from taking
- Authorized approver roles
- Information each request must include
- Escalation path for exceptions
- Audit and retention requirements
- Review date and accountable owner
NIST recommends defining and differentiating roles and responsibilities for human-AI configurations and oversight. NIST: Generative AI Profile →
Put Clear Boundaries Around AI Actions
Start with one AI-enabled workflow. Identify where information becomes an action, define who can approve it, test rejection and shutdown paths, and observe how the process works before expanding autonomy.
PCC helps Bay Area businesses strengthen the identity, permissions, Microsoft 365 and cybersecurity foundations that responsible AI adoption depends on.