Skip to main content

SOC 2 and AI: Using Governance Evidence for Audit Attestation

SOC 2 does not certify that an AI model is “safe,” and it does not provide a special carve-out for AI governance. It evaluates whether your organization designed and operated controls aligned to the Trust Services Criteria, typically around security, availability, confidentiality, processing integrity, and sometimes privacy. That means AI systems have to be folded into the same control logic as the rest of the environment. Auditors do not just want an AI policy deck. They want evidence that access, change control, monitoring, review, and exception handling actually happened.

Keeptrusts is useful because it produces evidence where many AI programs are weakest: the point where a real request becomes a real decision. The gateway can enforce route controls, return blocks or escalations, and emit decision events that can later be exported and reviewed. That does not mean Keeptrusts itself is your SOC 2 report. It means it can supply part of the evidence trail needed to support attestation over the AI portion of your control environment.

Use this page when​

  • You are preparing SOC 2 evidence for AI-enabled workflows, internal copilots, or customer-facing AI features.
  • You need a repeatable evidence model for auditors reviewing AI governance under existing Trust Services Criteria.
  • You want to strengthen the connection between route controls and audit-ready artifacts.

Primary audience​

  • Primary: compliance leads, security teams, internal audit, platform engineers
  • Secondary: GRC owners, external audit coordinators, product security

The problem​

AI programs often have a documentation surplus and an evidence deficit. Teams can describe their governance principles, provider review process, and incident-response expectations in detail, yet struggle to produce a clean package that shows how those controls were applied to actual model traffic. For SOC 2, that gap matters. Auditors are not evaluating aspirations. They are evaluating whether controls were designed appropriately and operated consistently over time.

The evidence problem shows up in several ways. Security controls become hard to prove when there is no clear record of blocked or escalated requests. Confidentiality controls become hard to prove when provider posture is not tied to runtime routing. Processing-integrity controls become hard to prove when there are no repeatable scenario tests for risky prompts or responses. Change-management controls become hard to prove when nobody can show which config governed a specific review period.

This is why AI teams sometimes create separate “AI audit” projects that duplicate work already required by SOC 2. A better approach is to make the AI route itself generate evidence that fits into the existing control environment. That keeps AI governance attached to the same discipline as other production systems.

The solution​

Use Keeptrusts as the runtime evidence layer for the AI portion of your SOC 2 narrative. Route controls can support several criteria families without pretending to satisfy them all on their own.

For security and confidentiality, prompt-injection, pii-detector, and data-routing-policy help show that risky input is screened and provider selection is constrained by declared handling posture. For processing integrity, financial-compliance, healthcare-compliance, or human-oversight can demonstrate that higher-risk outputs do not automatically flow to users without control logic. For monitoring and review, decision events and export jobs give you a consistent way to show what happened over a test period or audit window.

The strongest pattern is to combine route configuration with recurring validation. A tested configuration, visible decision events, and exported evidence together tell a better audit story than any one of those artifacts alone. The auditor sees not just that a control exists, but that it was configured, exercised, and monitored.

Implementation​

Use a short evidence loop that can be rerun every review period.

kt policy lint --file policy-config.yaml
kt policy test --json
kt gateway run --listen 0.0.0.0:41002 --policy-config policy-config.yaml
kt events tail --since 24h --json
kt export-jobs create --since 90d --format json --wait

That sequence covers the main audit questions. Was the config valid? Were expected scenarios tested? Did governed traffic produce decision records? Can the organization export a clean evidence package for the requested window? Pair those outputs with change-approval records, owner assignments, and the surrounding incident-response or vendor-management evidence already collected in your SOC 2 program.

The most helpful operational references are Pass Compliance Audits, kt export-jobs, Export Evidence for a Review, Reviewing Alerts and Evidence, and Configuration-First Workflow. Those pages give you the documented interfaces for the artifacts auditors usually ask to inspect.

Results and impact​

Organizations that adopt this model usually see audit preparation become more predictable. Instead of scrambling to explain AI with screenshots, ticket excerpts, and model-console exports, they can start from route-level evidence and then attach the broader control record around it.

There is also a quality benefit. When a team knows that route changes, policy tests, and event exports will become audit artifacts, it becomes much harder to let AI governance drift into informal exceptions and undocumented shortcuts.

Most importantly, this approach keeps AI inside the same assurance discipline as the rest of the product. That is generally what auditors and buyers want to see: not a special AI theater program, but operational control continuity.

Key takeaways​

  • SOC 2 does not certify AI behavior by itself; it evaluates whether your controls operated as designed.
  • AI programs need evidence, not just policy statements, if they want a credible audit story.
  • Keeptrusts can support that story through route controls, policy tests, decision events, and evidence exports.
  • The best evidence package combines configuration, test results, live event history, and review artifacts from the same period.
  • Keeptrusts contributes to attestation readiness, but it does not replace the broader SOC 2 control environment.

Next steps​