AI Agent Control Template

This template is used to assess, design, approve, monitor, and contain AI agents.

An AI agent is an AI system that can pursue a goal, plan steps, use tools, call APIs, trigger workflows, retrieve data, or perform actions with some level of autonomy.

Agentic AI requires stronger control than passive AI because it can move from generating output to executing actions.


1. Agent Information

Agent Name

[Enter agent name]

Agent ID

[Enter agent ID]

Date

[Enter date]

Agent Status

Select one:

[ ] Proposed
[ ] Under review
[ ] Approved
[ ] Pilot
[ ] Production
[ ] Restricted
[ ] Suspended
[ ] Under remediation
[ ] Retired
[ ] Rejected

Business Owner

Name:
Function:
Email:

Technical Owner

Name:
Function:
Email:

Agent Owner

Name:
Function:
Email:

Incident Contact

Name:
Function:
Email:
Escalation path:

2. Agent Purpose

Business Purpose

[Describe why this agent is needed and what business outcome it supports]

Agent Goal

[Describe the goal the agent is designed to pursue]

Approved Use Cases

[List approved use cases for this agent]

Prohibited Use Cases

[List prohibited or restricted uses]

User Population

Select all that apply:

[ ] Employees
[ ] Contractors
[ ] Developers
[ ] Security team
[ ] Customer support team
[ ] HR team
[ ] Finance team
[ ] Legal or compliance team
[ ] Business operations team
[ ] Customers
[ ] Suppliers
[ ] Partners
[ ] Public users
[ ] Other

3. Agent Autonomy

Autonomy Level

Select one:

[ ] Level 0: AI generates text only
[ ] Level 1: AI suggests actions but cannot execute
[ ] Level 2: AI prepares drafts but human executes
[ ] Level 3: AI requests actions but approval is required
[ ] Level 4: AI executes bounded low-risk actions
[ ] Level 5: AI executes high-impact actions under strict controls

Autonomy Description

[Describe what the agent can do independently and where human involvement is required]

Human Review Model

Select one:

[ ] Human-in-the-loop
[ ] Human-on-the-loop
[ ] Human-over-the-loop
[ ] Exception-based review
[ ] Not yet defined

Autonomy Limits

[Describe limits on planning, tool use, action execution, workflow scope, session length, or decision authority]

4. Agent Identity and Authority

Agent Identity Model

Select all that apply:

[ ] Direct user identity
[ ] Delegated user authority
[ ] Service identity
[ ] Application identity
[ ] Agent identity
[ ] Vendor-managed identity
[ ] Hybrid identity
[ ] Unknown

Agent Identity Details

Agent identity:
Service account:
Application identity:
Vendor identity:
Other:

Delegated Authority

[Describe whose authority the agent acts under, what is delegated, and under what conditions]

Access Scope

[Describe systems, repositories, APIs, workflows, tools, and data the agent can access]

Access Approval

Approver:
Date approved:
Approval reference:

Access Review Frequency

Select one:

[ ] Monthly
[ ] Quarterly
[ ] Semi-annually
[ ] Annually
[ ] At material change
[ ] Other

Revocation Path

[Describe how agent identity, delegated authority, tool access, API access, or service credentials can be revoked]

5. Data Boundary

Data Sources Accessible to Agent

Data Source Classification Owner Access Type Approved?
[Source] [Classification] [Owner] [Read/Write/Retrieve] [Yes/No/Pending]

Data Classes Accessible

Select all that apply:

[ ] Public
[ ] Internal
[ ] Confidential
[ ] Restricted
[ ] Regulated
[ ] Personal data
[ ] Customer data
[ ] Employee data
[ ] Financial data
[ ] Legal or privileged data
[ ] Security-sensitive data
[ ] Source code
[ ] Secrets or credentials
[ ] Production data
[ ] Other sensitive data

Retrieval Boundaries

[Describe what the agent can and cannot retrieve]

Data Restrictions

[Describe restricted data, prohibited data, masking, minimization, retention, or reuse restrictions]

Vendor Processing

[Describe whether agent inputs, context, outputs, logs, or tool calls are processed by a vendor]

6. Prompt and Input Control

Input Sources

Select all that apply:

[ ] User prompt
[ ] System prompt
[ ] Retrieved document
[ ] Email
[ ] Ticket
[ ] Chat message
[ ] Customer submission
[ ] API payload
[ ] Tool response
[ ] Web content
[ ] Code
[ ] Log data
[ ] Workflow state
[ ] Conversation history
[ ] Memory
[ ] Other

External or Untrusted Inputs

[Describe external or untrusted inputs processed by the agent]

Prompt Injection Risk

Select one:

[ ] Low
[ ] Moderate
[ ] High
[ ] Critical
[ ] Unknown

Prompt Injection Controls

[Describe controls to prevent untrusted content from overriding system instructions or causing unsafe tool use]

System Prompt Protection

[Describe system prompt ownership, versioning, access control, testing, and rollback]

Context Isolation

[Describe how context is isolated across users, sessions, tenants, repositories, trust levels, and data classifications]

7. Tool Inventory

Tools Available to Agent

Tool / API / Workflow Owner Action Type Risk Level Approval Required? Logging Available?
[Tool] [Owner] [Read/Draft/Write/Action/Admin] [Low/Medium/High/Critical] [Yes/No] [Yes/No]

Tool Permissioning

[Describe which tools are allowed, restricted, prohibited, or conditionally available]

Prohibited Tools

[List tools, APIs, workflows, or actions the agent must not use]

Tool Access Review

Review frequency:
Reviewer:
Last reviewed:
Next review:

8. Action Classification

Agent Actions

Action Description Risk Level Reversible? Approval Required?
[Action] [Description] [Low/Medium/High/Critical] [Yes/No/Partial] [Yes/No]

High-Risk Actions

Select all that apply:

[ ] Customer-impacting action
[ ] Employee-impacting action
[ ] Financial action
[ ] Legal or compliance action
[ ] Security action
[ ] Production change
[ ] Access grant or revocation
[ ] Record modification
[ ] External communication
[ ] Regulated workflow action
[ ] Irreversible or hard-to-reverse action
[ ] Other

Action Notes

[Describe action risks, reversibility, downstream impact, and approval expectations]

9. Approval Gates

Approval Required Before Execution?

[ ] No
[ ] Yes
[ ] Conditional
[ ] Unknown

Approval Gate Summary

Action / Tool Approval Trigger Approver Evidence Required Expiry
[Action] [Trigger] [Approver] [Evidence] [Expiry]

Approval Conditions

[Describe thresholds, conditions, or exceptions that require approval]

Approval Bypass Prevention

[Describe controls that prevent the agent from bypassing approval]

10. Execution Boundaries

Boundaries Applied

Select all that apply:

[ ] Tool boundary
[ ] Data boundary
[ ] System boundary
[ ] Environment boundary
[ ] User boundary
[ ] Role boundary
[ ] Workflow boundary
[ ] Action boundary
[ ] Time boundary
[ ] Rate boundary
[ ] Amount boundary
[ ] Geography boundary
[ ] Customer segment boundary
[ ] Other

Boundary Details

[Describe exact execution boundaries]

Boundary Enforcement

[Describe how boundaries are technically or procedurally enforced]

Boundary Violation Response

[Describe what happens if the agent attempts to exceed boundaries]

11. Blast-Radius Limits

Limits Applied

Limit Type Limit Value Enforcement Method Owner
Number of records [Limit] [Method] [Owner]
Transaction value [Limit] [Method] [Owner]
Number of users affected [Limit] [Method] [Owner]
API calls [Limit] [Method] [Owner]
Time window [Limit] [Method] [Owner]
Systems affected [Limit] [Method] [Owner]
Environment [Limit] [Method] [Owner]
Other [Limit] [Method] [Owner]

Blast-Radius Notes

[Describe maximum potential impact if the agent fails or is misused]

12. Monitoring and Logging

Logs Required

Select all that apply:

[ ] Agent session logs
[ ] User prompt logs
[ ] Prompt metadata
[ ] Retrieved context logs
[ ] Tool call logs
[ ] Action logs
[ ] Approval logs
[ ] Policy decision logs
[ ] Boundary violation logs
[ ] Error logs
[ ] Output logs
[ ] Monitoring alerts
[ ] Incident records

Required Log Fields

Select all that apply:

[ ] User identity
[ ] Agent identity
[ ] Delegated authority
[ ] Session ID
[ ] Goal or request
[ ] Plan or reasoning summary
[ ] Tool selected
[ ] Tool input
[ ] Tool output
[ ] Action requested
[ ] Action executed
[ ] Approval status
[ ] Result
[ ] Exception
[ ] Timestamp
[ ] Policy decision
[ ] Error or failure

Log Location

[Describe where logs are stored]

Log Retention

[Describe retention period and access restrictions]

Monitoring Rules

[Describe monitoring rules for abnormal tool use, action volume, policy violations, failed approvals, prompt injection, or unusual behavior]

13. Kill Switch and Containment

Kill Switch Available?

[ ] No
[ ] Yes
[ ] Partial
[ ] Unknown

Kill Switch Level

Select all that apply:

[ ] Agent level
[ ] Tool level
[ ] API level
[ ] Workflow level
[ ] Identity level
[ ] Data source level
[ ] Vendor feature level
[ ] User group level
[ ] Action category level
[ ] Environment level

Kill Switch Owner

Name:
Function:
Email:

Activation Conditions

[Describe when the kill switch should be activated]

Activation Process

[Describe how the kill switch is activated]

Expected Time to Disable

[Enter expected time]

Restart Conditions

[Describe approval and testing required before restart]

Last Kill Switch Test

Date:
Result:
Evidence:

14. Rollback and Recovery

Rollback Available?

[ ] No
[ ] Yes
[ ] Partial
[ ] Unknown

Rollback / Recovery Summary

Action Type Recovery Method Owner Tested? Evidence
[Action] [Method] [Owner] [Yes/No] [Evidence]

Compensation Required If Rollback Is Not Possible?

[ ] No
[ ] Yes
[ ] Unknown

Compensation Method

[Describe compensation, correction, customer notification, manual remediation, or record amendment]

Recovery Notes

[Describe recovery constraints, dependencies, and approval requirements]

15. Assurance and Testing

Required Agent Tests

Select all that apply:

[ ] Agent identity test
[ ] Tool permission test
[ ] Unauthorized tool call test
[ ] Prompt injection to action test
[ ] Approval bypass test
[ ] Action boundary test
[ ] Blast-radius limit test
[ ] Logging completeness test
[ ] Kill switch test
[ ] Rollback test
[ ] Incident tabletop
[ ] Regression test

Test Results Summary

Test Result Evidence Finding
[Test] [Pass/Fail/Partial/Not Run] [Evidence] [Finding]

Open Findings

Finding ID Finding Severity Owner Due Date Status
[Finding ID] [Finding] [Low/Medium/High/Critical] [Owner] [Date] [Status]

16. Incident Scenarios

Relevant Agent Incident Scenarios

Select all that apply:

[ ] Agent performs unauthorized action
[ ] Agent calls unauthorized tool
[ ] Agent follows prompt injection
[ ] Agent leaks data through tool use
[ ] Agent modifies incorrect record
[ ] Agent sends unauthorized communication
[ ] Agent loops or retries excessively
[ ] Agent exceeds blast-radius limit
[ ] Agent bypasses approval
[ ] Agent causes customer impact
[ ] Agent causes production impact
[ ] Agent causes security impact
[ ] Agent cannot be stopped quickly
[ ] Agent action cannot be rolled back

Incident Response Path

[Describe escalation, containment, evidence preservation, investigation, recovery, and communication path]

17. Risk Assessment

Agent Risk Rating

Select one:

[ ] Low
[ ] Medium
[ ] High
[ ] Critical

Risk Drivers

Select all that apply:

[ ] Sensitive data access
[ ] Regulated data access
[ ] Tool/action capability
[ ] High autonomy
[ ] Customer impact
[ ] Employee impact
[ ] Financial impact
[ ] Security impact
[ ] Production impact
[ ] External communication
[ ] Vendor dependency
[ ] Weak logs
[ ] Weak rollback
[ ] Weak kill switch
[ ] Prompt injection exposure
[ ] Approval bypass risk

Risk Notes

[Describe agent risk and residual risk]

18. Approval Decision

Decision

Select one:

[ ] Approved
[ ] Approved with conditions
[ ] Approved for pilot only
[ ] Requires remediation
[ ] Requires additional testing
[ ] Requires exception approval
[ ] Rejected
[ ] Deferred

Conditions

[List required conditions before agent approval, production use, scaling, or restart]

Exceptions Required?

[ ] No
[ ] Yes
[ ] Unknown

Residual Risk Acceptance Required?

[ ] No
[ ] Yes
[ ] Unknown

19. Approval Record

Business Owner Approval

Name:
Decision:
Date:
Notes:

Technical Owner Approval

Name:
Decision:
Date:
Notes:

Security / Architecture Approval

Name or forum:
Decision:
Date:
Notes:

Risk / Governance Approval

Name or forum:
Decision:
Date:
Notes:

Incident Response Approval, If Required

Name or forum:
Decision:
Date:
Notes:

20. Summary

Agent:
Use case:
Owner:
Autonomy level:
Identity model:
Data accessed:
Tools available:
High-risk actions:
Approval gates:
Blast-radius limits:
Kill switch:
Rollback:
Logging:
Testing status:
Risk rating:
Approval status:
Next review date: