SOC 2 Control Implementation Guide

AI / LLM Governance

Approved Use Cases and Human Review for SOC 2

AI/LLM outputs affecting customer-facing communications, incident decisions, security recommendations, or automated actions require defined human review.

Use this guide to put the control into operation, decide what records to retain, and check that an auditor can trace the evidence back to the work your team performed.

Maintained by GreenHat Security · Reviewed August 21, 2026

What this control should accomplish

AI use is limited to approved purposes, and consequential outputs or actions receive a defined human decision before they affect customers or operations.

First SOC 2 program

A credible starting point

List approved use cases and place a named reviewer in the workflow before customer communications, security decisions, or external actions are released.

As the company scales

Make it repeatable

Use risk-tiered review thresholds, reviewer queues, escalation rules, measured override rates, and quality sampling for high-volume workflows.

How to implement Approved Use Cases and Human Review

  1. 1

    Approve the use case

    Document intended users, allowed decisions, prohibited uses, input data, output destination, and accountable owner before launch.

    You should end up with: Signed use-case decision

  2. 2

    Define the human decision

    State which outputs require review, reviewer qualifications, information presented, approval choices, and escalation conditions.

    You should end up with: Human-review procedure and decision criteria

  3. 3

    Enforce the review point

    Prevent covered output or actions from continuing until the designated reviewer records a decision.

    You should end up with: Workflow gate with attributable decision log

  4. 4

    Assess actual use

    Compare workflow activity with approved purposes and review approvals, rejections, overrides, and missed escalations.

    You should end up with: Use-case and reviewer performance review

Evidence to keep, and what it should prove

Build the evidence set in three layers: what defines the control, who approved or reviewed it, and what proves it operated. Collect operating records when the work happens so they remain dated, attributable, correctly scoped, and traceable to the underlying activity.

Before sharing, remove unrelated personal or customer data, never expose passwords, tokens, or secret values, preserve enough source context to authenticate the record, and use the secure exchange approved for the engagement.

Policy / design artifacts

Documents that define the control, its scope, ownership, and expected way of working.

human review requirements

  • Confirm what the record proves

    Consequential outputs have explicit review triggers, qualified reviewers, decision criteria, escalation routes, and enforced release gates.

  • Include this context

    Covered output or action

  • Include this context

    Trigger

  • Include this context

    Reviewer role

  • Include this context

    Decision criteria

  • Include this context

    Escalation path

  • Include this context

    Enforcement point

Weak evidence to avoid

A statement that humans stay in the loop without defining which outputs, who reviews, or how release is blocked.

Approval / review evidence

Records showing that an accountable person reviewed, approved, challenged, or accepted the work.

Use case approvals

  • Confirm what the record proves

    The AI workflow is operating for a defined purpose, audience, data scope, and decision boundary approved by an accountable owner.

  • Include this context

    Use-case ID

  • Include this context

    Workflow ID

  • Include this context

    Allowed and prohibited purposes

  • Include this context

    Data scope

  • Include this context

    Approver

  • Include this context

    Decision and date

Weak evidence to avoid

An email saying the AI feature is approved without its users, data, output destination, or prohibited uses.

reviewed output records

  • Confirm what the record proves

    Covered runtime outputs were actually reviewed and resulted in attributable approval, correction, rejection, or escalation before effect.

  • Include this context

    Execution or output ID

  • Include this context

    Workflow version

  • Include this context

    Reviewer

  • Include this context

    Review timestamp

  • Include this context

    Decision

  • Include this context

    Released or corrected result

Weak evidence to avoid

A folder of generated answers with no reviewer identity, decision, timestamp, or indication of what reached the customer.

Which records should you prepare for the audit?

Type 1

Evidence at the as-of date

Retain current approved use cases and human-review rules, the latest scheduled annual review and its disposition, plus one recent covered output showing the enforced review and final decision on the selected date.

Type 2

Evidence across the review period

Export every runtime output or proposed action during the period that met a human-review trigger, including approvals, corrections, rejections, escalations, timeouts, cancellations, and any bypass attempt; include every use-case or review-rule change and every scheduled annual use-case and human-review effectiveness review, including no-change completions, missed reviews, and rescheduled reviews.

Completeness check

Reconcile review-queue records to all runtime outputs and actions matching configured triggers, then compare completed approvals with customer-facing messages and executed actions. Separately reconcile every annual review due date from the schedule to a completed no-change or change decision, a documented missed review, or a rescheduled review that preserves the original and new due dates.

Build the record set from

  • Workflow runtime logs
  • Human-review queue
  • Approval service
  • Customer communication or action system
  • Use-case register
  • Workflow configuration history
  • Annual review schedule and review records

Keep these fields for each record

  • Execution and output IDs
  • Workflow version
  • Use case
  • Trigger
  • Reviewer
  • Runtime decision and timestamp
  • Final disposition
  • Bypass or escalation flag
  • Annual review due date
  • Annual review status
  • Completion or rescheduled date
  • Review decision and follow-up

How an auditor may test this control

Use this checklist to prepare for procedures an auditor may perform. The exact steps and sample selection depend on your engagement scope and the service auditor's professional judgment.

  • Confirm the intended control outcome

    Determine whether the control is designed to achieve this result: AI use is limited to approved purposes, and consequential outputs or actions receive a defined human decision before they affect customers or operations.

  • Confirm ownership and operating cadence

    Compare the documented owner with the intended role (AI Workflow Owner / CISO / Engineering), then compare dated records with the stated cadence: Per workflow/change; review at least annually.

  • Establish the complete audit record set

    Reconcile review-queue records to all runtime outputs and actions matching configured triggers, then compare completed approvals with customer-facing messages and executed actions. Separately reconcile every annual review due date from the schedule to a completed no-change or change decision, a documented missed review, or a rescheduled review that preserves the original and new due dates.

  • Prepare the as-of-date evidence for a Type 1 engagement

    Retain current approved use cases and human-review rules, the latest scheduled annual review and its disposition, plus one recent covered output showing the enforced review and final decision on the selected date.

  • Prepare period evidence for a Type 2 engagement

    Export every runtime output or proposed action during the period that met a human-review trigger, including approvals, corrections, rejections, escalations, timeouts, cancellations, and any bypass attempt; include every use-case or review-rule change and every scheduled annual use-case and human-review effectiveness review, including no-change completions, missed reviews, and rescheduled reviews.

  • Inspect the policy / design artifacts

    Documents that define the control, its scope, ownership, and expected way of working.

    • Inspect human review requirements

      For each selected record, confirm it demonstrates Consequential outputs have explicit review triggers, qualified reviewers, decision criteria, escalation routes, and enforced release gates.

      • Covered output or action
      • Trigger
      • Reviewer role
      • Decision criteria
      • Escalation path
      • Enforcement point
  • Inspect the approval / review evidence

    Records showing that an accountable person reviewed, approved, challenged, or accepted the work.

    • Inspect Use case approvals

      For each selected record, confirm it demonstrates The AI workflow is operating for a defined purpose, audience, data scope, and decision boundary approved by an accountable owner.

      • Use-case ID
      • Workflow ID
      • Allowed and prohibited purposes
      • Data scope
      • Approver
      • Decision and date
    • Inspect reviewed output records

      For each selected record, confirm it demonstrates Covered runtime outputs were actually reviewed and resulted in attributable approval, correction, rejection, or escalation before effect.

      • Execution or output ID
      • Workflow version
      • Reviewer
      • Review timestamp
      • Decision
      • Released or corrected result
  • Trace the control from design to operation

    Use the categories that apply to this control: connect any policy or design artifact to its approval or review record, then trace a selected operating record through execution, result, and any exception or remediation.

Common implementation and evidence gaps

  • Human review is described in policy but the workflow can bypass it.
  • Reviewers see the generated answer without supporting context or evidence.
  • The approved use case expands quietly through prompt or product changes.
  • Approval logs do not identify the reviewer or final decision.

Before you call this control ready

  • Can a covered output reach a customer without a recorded decision?
  • Do reviewers know when to reject or escalate an output?
  • Can recent activity be tied to an approved use case and owner?
  • Are rejection and override patterns reviewed for workflow changes?

Trust Services Criteria references

These identifiers help you navigate related Trust Services Criteria. They do not reproduce the criteria or prove that this control fully addresses them in your environment.

  • CC5.1
  • CC8.1
  • P8.1

Confirm final scope, mappings, and testing expectations with your service auditor. SOC 2® is an AICPA trademark; GreenHat Security is not affiliated with or endorsed by AICPA.