SOC 2 doesn't mention AI coding tools. It doesn't need to. The change management criterion (CC8.1 in the Trust Services Criteria) already expects that changes to production systems are authorized, tested, approved and documented. When AI writes a growing share of those changes, auditors are starting to ask how that still holds.
This article is general information, not audit or legal advice. Your auditor's expectations are what count.
The questions that come up
Based on what security and compliance teams describe, the questions tend to be some version of:
- Who reviewed this change? A named person, not "the AI reviewed it."
- What was tested, and on which version? Evidence that the tests ran on the code that shipped.
- Was the change authorized? It traces back to a ticket, request or plan.
- Can a single actor push to production alone? Separation between the author (human or agent) and the approver.
- Can you show this for a sample? Auditors pick changes from the period and ask for the record.
None of these are new. What's new is volume: AI agents produce far more changes, and the evidence for each has to come from somewhere.
Where teams get stuck
- Approvals that don't match what shipped. A reviewer approved one commit, the agent pushed a fix, and the merge happened without a fresh look.
- "AI reviewed it" as the only review. An AI review is useful input, but it isn't a human approval.
- Evidence spread across tools. Ticket in one place, CI logs in another, approval in the pull request, and nothing linking them.
- Agent permissions. Agents with tokens that can merge or push to protected branches remove the separation auditors look for.
What good looks like
- Every change links to its authorization (ticket or plan).
- Tests run on the exact commit that merges, and the result is kept.
- A named person approves, and approvals reset on new commits.
- Agents can propose changes but can't approve or merge their own work.
- One record per change brings this together, and you can export it for any sample.
How AI CodeOps helps
AI CodeOps was built around this record. AI agents plan, build, review and test changes on your repository, and a pull request only opens when your tests pass on that exact commit, the AI review of the real diff has no open high or critical findings, and a person approves. Policies control what each agent can do, and agents can't approve their own changes.
Every change gets a signed evidence pack: what was requested, what changed, which tests ran on which commit, what the review found and who approved it. When an auditor picks a sample, the record already exists.
Free during the open beta: https://aicodeops.dev
Top comments (0)