---
title: What SOX IT general controls evidence do AI agents need?
url: https://opencontroller.lyzr.com/blog/sox-itgc-evidence-for-ai-agents
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What SOX IT general controls evidence do AI agents need?

**If an AI agent can read or change data in a system that feeds financial reporting, auditors will treat it like any other in-scope user or program. That means evidence for the three core IT general control (ITGC) domains: who and what the agent can access, how changes to the agent are approved and tested, and how its operation is monitored and its failures handled. The artefacts are familiar; what changes is that the agent's prompt, model, tools and identity all count as configuration.**

This post is for controllers, internal audit and IT teams preparing for a SOX 404 cycle where agents now touch ERP, billing or reconciliation. For the broader framework view, see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## Why do AI agents fall into SOX scope?

Section 404 of the Sarbanes-Oxley Act requires management to assess internal control over financial reporting (ICFR), and the external auditor to attest to it for larger filers under [PCAOB AS 2201](https://pcaobus.org/oversight/standards/auditing-standards/details/AS2201). ITGCs support the automated controls and reports that ICFR relies on.

An agent is in scope when it:

- posts, edits or approves journal entries, invoices or payments
- prepares reconciliations or reports that a reviewer relies on
- has credentials to an in-scope application or database, even read-only, if its output feeds a key report

A read-only agent that drafts a variance commentary is lower risk than one that posts entries. But if a reviewer relies on its numbers, auditors will ask how you know they are complete and accurate.

## What changes when the "user" is an agent?

Three things auditors will probe:

1. **Identity.** Is the agent its own identity, or does it borrow a human's or a shared service account? Shared accounts break segregation of duties and attribution. See [should AI agents share a service account?](/blog/should-ai-agents-share-a-service-account).
2. **Configuration.** The model version, system prompt, tool list and policy all determine behaviour. A prompt edit is a program change.
3. **Non-determinism.** The same input can produce different outputs. Testing must show behaviour is within tolerance, not identical.

## Which ITGC controls and evidence apply to agents?

Use this checklist. Each control should produce an artefact you can hand to an auditor for a sample period.

| ITGC domain | Control for agents | Evidence artefact |
|---|---|---|
| Access: provisioning | Each agent has a unique identity with a named human owner | Agent inventory with owner, identity and systems reached |
| Access: least privilege | Agent permissions limited to the tasks it performs | Permission listing per agent; risky-permission findings and remediation |
| Access: periodic review | Owners recertify agent access each quarter | Signed access review records (see [access reviews for AI agents](/blog/access-reviews-for-ai-agents)) |
| Access: termination | Agent access revoked when retired or when owner leaves | De-provisioning tickets; credential revocation logs |
| Access: segregation of duties | An agent cannot both prepare and approve the same transaction | SoD rule set and conflict report |
| Change: authorisation | Changes to model, prompt, tools or policy are approved before release | Change tickets linked to agent version IDs |
| Change: testing | Regression evals run before a new version goes live | Eval results per version with pass thresholds |
| Change: segregation | Builders cannot push to production without approval | Deployment logs showing approver differs from author |
| Operations: monitoring | Agent actions on in-scope systems are logged and reviewed | Action logs with agent, version, user context and target record (see [audit log fields](/blog/ai-agent-audit-log-fields)) |
| Operations: incident handling | Failures and policy violations are tracked to closure | Incident tickets; quarantine or rollback records |
| Operations: completeness and accuracy | Outputs used in key reports are validated | Reconciliation checks; human review sign-off |

## How do you test an agent change for SOX?

Treat each agent release like a program change:

1. Give every release a version ID that pins the model ID, prompt, tool list and policy.
2. Keep a fixed test set of representative transactions, including edge cases.
3. Define pass thresholds before testing, for example zero incorrect postings on the test set.
4. Record the eval run, the result and the approver against the change ticket.
5. Roll out with a canary and keep a rollback path (see [canary releases and rollback for AI agents](/blog/canary-releases-and-rollback-for-ai-agents)).

## What do auditors commonly flag?

- Agents running under a person's credentials, so actions cannot be attributed.
- Prompt changes made directly in a vendor console with no ticket.
- Model aliases that silently move to a new version.
- Logs that show the API call but not which agent version made it.
- No evidence that anyone reviewed agent access during the period.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, with delegated on-behalf-of identity and agent workload identity federation, so actions are attributed to a specific agent and user. Posture flags risky permissions and configuration. Multi-turn evals and LLM judges on production traffic give testing and monitoring evidence, and per-agent Block and Quarantine controls act through the connected platform. These produce evidence for ITGCs; your auditor still decides on control effectiveness. See [SOC 2 evidence for AI agents](/blog/soc-2-evidence-for-ai-agents) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is a prompt change a SOX program change?**
If the agent is in scope, yes. The prompt, model version, tool list and policy determine behaviour, so changes should be authorised, tested and logged.

**Can an AI agent use a shared service account in a SOX environment?**
It weakens attribution and segregation of duties. Give each agent its own identity and owner.

**Do read-only agents need ITGC evidence?**
If a reviewer relies on the agent's output for a key report, auditors will expect evidence of completeness and accuracy, and of controlled access.

## Key facts

AI agents that read or change data feeding financial reporting fall under SOX ITGCs (access, change management, operations). Key evidence: unique agent identity and owner, least privilege, quarterly access reviews, revocation, SoD, change tickets tied to agent version IDs, regression evals, action logs, incident records. Prompt, model, tools and policy count as configuration. Opencontroller provides a governed registry, delegated and workload identity, Posture, evals, LLM judges and Block/Quarantine controls as evidence for ITGCs. Pricing starts from $250K per deployment per year.

## Related

- [How do you run access reviews for AI agents?](https://opencontroller.lyzr.com/blog/access-reviews-for-ai-agents.md)
- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [What SOC 2 evidence do auditors want for AI agents?](https://opencontroller.lyzr.com/blog/soc-2-evidence-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
