Guides

What does a SOC 2 audit expect from your AI usage?

SOC 2 doesn't have an AI section, but Type II auditors sample evidence, and LLM calls are activity your controls have to cover. What auditors ask about AI usage, and what evidence answers it.

SOC 2 has no AI chapter. That doesn't help you, because a Type II audit samples evidence of controls operating, and if your product or your internal teams call LLMs, that activity is inside your system boundary and your controls have to cover it. Auditors have caught up to this faster than most vendors have.

The questions that now show up

• Data handling: customer data leaves your environment in prompts. What screens it before it goes, and does the screening fail open or closed when it errors?
• Access and attribution: who can make AI calls, with which credentials, and can a given call be attributed to a person or service rather than a shared key?
• Logging and monitoring: is AI activity recorded like other system activity, and can the log itself be trusted, or can anyone with database access rewrite history?
• Vendor management: model providers are subservice organizations. Which ones do you use, under what terms, and does your inventory match reality, including the endpoints individual developers wired up without telling anyone?

Why tamper-evidence upgrades the answer

Most of those questions can be answered adequately with ordinary discipline. The one that separates a smooth audit from a finding is the third: an auditor who asks "how do you know this log is complete and unmodified" is asking a question ordinary application logging cannot answer. A hash-chained record with a signed export can: any edit, deletion, or reordering of past records breaks every subsequent hash, and the auditor checks it with a standalone tool rather than your word. That is the difference between evidence the auditor samples and a story the auditor decides whether to believe. The mechanism, in detail, is on the trust page.

One layer instead of four controls

This is what RemKey is for: one governed gateway in front of all LLM traffic that collapses the control set. Screening (fail-closed), attribution (per-key, per-end-user), tamper-evident logging on a hash-chained, Ed25519-signed ledger, and a live inventory of which models are actually in use, all produced by the same layer that routes the traffic, all landing in one evidence trail. One base URL to adopt, one export to hand the auditor, verified offline without trusting us. And because we hold ourselves to the same standard: RemKey's own SOC 2 control environment is documented and our current status is stated plainly on the trust page, we will show you the real state in the first call rather than imply a report we don't hold.