How to Document AI Workflows for a SOC 2 Type II Audit
A practical guide to the records that prove your AI controls worked.

A SOC 2 Type II audit checks whether your security and data controls actually worked over a set period of time. For an AI workflow, that means showing how you kept the system under control. Models, settings, data sources, and permissions all change. You need records that cover all of that.
Start by identifying which systems are being audited. Work out what you have promised customers about those systems. Then decide which of the AICPA Trust Services Criteria (the official checklist of control categories published by the accounting standards body) apply to you. Match your records to those controls. Agree with the auditor what evidence is needed. A documentation checklist on its own does not prove your controls worked.
What is different about auditing an AI workflow
Standard change management records what changed, who approved it, how it was tested, and when it went live. AI workflows still need all of that. But they also need records for actions the AI takes on its own at runtime, within a scope that was approved in advance.
Three things deserve extra attention.
First, AI outputs can vary even when the input is identical. Testing should check whether the workflow stays within defined quality and risk limits across a range of cases. It should not just confirm that the wording matches a fixed answer. Keep the test cases, the pass or fail criteria, the results, and the action taken when something fails.
Second, some AI work starts automatically on a schedule rather than because a person clicked a button. Record the identity of the agent or service doing the work. Record the scope it was approved to operate within, who owns it, and what permissions it checked before each action. Where a person did start or approve the work, link their identity to that run. An automated action can be properly authorized without a person approving every individual step.
Third, an external model provider may change something in the underlying system without you changing any of your own code. Record the model name and version information the provider gives you. Record your relevant settings and any changes in behavior you observe. If the provider does not give you an exact version number, write that down. Also record how you monitor or test for unexpected behavior changes.
Unapproved AI tools create another risk. Company data may reach services that have never been assessed. Keep an inventory of approved tools and providers. Define what data may be used with them. Record how you spot and resolve exceptions. Your control design should reflect how employees and systems actually use AI.
What your AI audit trail should show
A timestamp alone rarely explains an AI-assisted decision. A useful record connects the original request or scheduled trigger to the authority under which the workflow ran. It also covers the information the workflow used, the actions it took, and the result. Protect that record against unauthorized changes. Make sure you can retrieve it when needed for a control test.
Decide what fields to keep for each workflow. Useful ones include:
The workflow and run identifiers, timestamps, purpose, and accountable owner.
The requesting person or scheduled trigger, the agent or service that ran, the approved scope, and the permissions that applied.
Relevant inputs and source references, the version of the prompt or configuration, model identifiers, and tool versions.
Available reasoning behind decisions, validation results, policy checks, approvals, exceptions, and escalation records.
Tool calls, significant changes made, outputs, failures, corrections, and measured results.
Choose what to keep based on the control objective and how sensitive the data is. It may be better to keep a protected reference or a redacted record than to copy a full prompt, document, or API response (a data request sent between systems) into every log. Your evidence design should still let an authorized reviewer retrieve enough information to assess the control.
A decision record — a central record organized around a common run identifier — can connect the version and inputs used, the rule or threshold applied, the approval in force, and the observed result. It is one way to do this, not a substitute for defining the control or checking that the record is complete and accurate.
Service accounts (automated system identities, as opposed to individual user accounts) need a named owner and controlled permissions. Avoid credentials shared across unrelated workflows. Shared credentials prevent reviewers from identifying which workflow acted. Distinguish the identity that carried out an action from the person or role that approved its scope.
Organizing your evidence across the five control categories
The AICPA framework covers five categories: security, availability, processing integrity, confidentiality, and privacy. Use the categories that fall within your audit scope. The points below are control-design considerations. They are not a claim that every AI workflow needs the same controls.
Security. Document access control, privileged identities, approved tool use, change management, and your assessment of relevant model and tool providers. Include evidence that access restrictions work as intended. Show that inappropriate access is investigated. Hosting a model privately may support a chosen control design, but it does not replace access management, monitoring, or vendor assessment.
Processing integrity. Define what complete, accurate, and authorized processing means for the workflow. Keep relevant validation results, acceptance criteria, records of exceptions, and checks run before significant downstream changes. Choose monitoring frequency and escalation thresholds that fit the workflow's risks. A significant change in observed behavior should trigger a documented response.
Confidentiality and privacy. Document which data enters prompts, retrieval systems, tools, and logs. Record who can access it and how it is protected and eventually deleted. Apply minimization, redaction, access restrictions, or other safeguards to the evidence store itself. Logging sensitive information creates another copy that must be managed. The AICPA's privacy considerations for SOC 2 examinations give more context when privacy is in scope.
Availability. Record what happens when a model endpoint, tool, or connected system is unavailable. Test fallback, retry, and recovery behavior. Check whether retries could accidentally repeat an action. Keep the dependency information, incident records, and recovery-test evidence relevant to your service commitments.
How long to keep records, and what other rules apply
Build a retention schedule by record category rather than assigning one time period to every AI log. Operational logs, approval records, personal data, and audit workpapers serve different purposes and have different owners. Identify the legal, contractual, and business requirements that apply before you set retention or deletion rules.
For each category, record its purpose, owner, storage location, access rules, retention period, and disposal process. Include how legal holds or other valid exceptions to deletion are handled. Check that archived records can actually be retrieved when needed. Make sure deletion rules cover copies in connected systems.
Document where model processing and log storage occur. Include relevant provider arrangements and any transfer of data across borders. Review this when providers, data categories, or deployment locations change. A SOC 2 report should not be treated as confirmation that every separate legal obligation has been met.
Ownership and accountability
Assign responsibility for model and tool selection, deployment, access management, validation, monitoring, and exception handling. The record should make clear who approved the workflow's operating scope. It should also show who can change, pause, or revoke it.
For recurring automated work, define permitted actions, data boundaries, quality standards, risk limits, and escalation conditions before deployment. A person approves the scope within limits they set. The system then checks current permissions when an action runs. That approved scope should not silently expand because the model selected a new tool or because an earlier action happened to succeed.
Keep scope changes and exceptions visible. Record what triggered an escalation, who decided the response, and whether the workflow was corrected or restricted. Review ownership and access when people change roles or leave the organization. These records connect machine actions to accountable human decisions.
Building your evidence across the whole review period
A screenshot can show a setting at one moment. Evidence of a recurring control also needs to show that the activity happened at the frequency you committed to. Depending on the control, that may mean dated reviews, execution logs, alerts, exception records, or test results. You do not necessarily need to collect every possible event continuously.
Maintain an evidence register that connects each control to its owner, frequency, expected records, and storage location. Useful records include the workflow and model inventory, production change approvals, validation results, scheduled access reviews, provider reassessments, and incident responses.
Monitor the evidence pipeline itself. Missing events, failed exports, or inaccessible archives can prevent a reviewer from following a run even when the workflow completed successfully. Record these gaps and their resolution. An alert log is useful evidence, but it does not by itself show that alerts were complete, investigated, or resolved.
Periodically pick a completed run and trace it end to end. Check the approved scope, relevant inputs and versions, checks during execution, downstream action, and observed result. Include exceptions and failures in these reviews. That exercise exposes disconnected records before they become hard to reconstruct during an audit.
How a knowledge graph can help with traceability
A knowledge graph — a database that records how things relate to each other — is one way to connect a workflow, its owner, model and tool versions, source evidence, approvals, actions, and outcomes. Those relationships can help a reviewer follow a decision across systems without manually matching separate log files.
Preserve source references, timestamps, and access controls as those relationships change. A graph should make missing evidence and conflicting records visible. It should not make an incomplete trail look complete. Its value is making the relevant records easier to retrieve and assess. The graph itself does not show that controls worked.
Start with one workflow. Confirm that an authorized reviewer can follow a real run from approved scope to outcome. Then extend the same ownership, versioning, and access rules to other workflows.
Sources and further reading
Primary framework references
Related vendor commentary
The following links offer implementation perspectives. They are not sources of authoritative SOC 2 requirements.


