Audit Légal + Conformité Assisté par IA

What's in this pack

This is the rig the in-house compliance officer, SOC2-prep lead, or privacy officer would assemble for an enterprise audit cycle — explicitly not the same problem a single lawyer faces when redlining one MSA. The audience here owns a recurring program: a risk register that gets updated quarterly, control mappings that have to survive an auditor walkthrough, evidence that has to be both collectable and tamper-evident, and a policy library that drifts the moment engineering ships something.

The whole stack is built around three principles that compliance work demands and consumer AI tools rarely deliver:

1. Document the AI's role, don't hide it. Auditors increasingly ask how AI was used in evidence collection or control testing. Every pick here produces a record an auditor can read.
2. Sensitive material never leaks to a vendor. Internal policies, draft risk assessments, and customer PII go through redaction or stay on infrastructure you control.
3. The AI proposes; the human decides. Materiality, risk acceptance, and any judgment that lands in an opinion letter remain a human responsibility. The tooling drafts, extracts, and maps — it does not sign off.

Nothing in this kit is legal advice. None of it replaces a qualified compliance professional, external auditor, or counsel. These are infrastructure picks for the mechanical layer of compliance work so the humans can spend their hours on judgment.

Install in this order (intake → policy gap → risk register → control map → audit log → report)

1. Claude Code Agent: Compliance Auditor — the orchestrator. A focused subagent profile that knows the language of GDPR, HIPAA, PCI DSS, SOC 2, and ISO 27001-family controls. Use it as the entry point: it asks what framework you're working against, what scope you're testing, and what evidence you've already collected. Treat its output as a senior analyst's first draft, not a final position.
2. Claude Code Agent: Compliance Auditor — Regulatory Checks — a sibling agent oriented to regulatory checks rather than internal framework gap analysis. Run the two in tandem: the first scopes the audit, the second pressure-tests specific regulatory exposure (cross-border data, sectoral rules, breach-notification timelines). Two agents catch what one misses.
3. Agent Governance Toolkit — Policy Guardrails for Agents — your policy gap analysis layer. Compliance teams are increasingly responsible for AI use across the business; this toolkit lets you write the guardrails (what models, what data classes, what regions) once and enforce them at the agent layer. Closes the most-asked SOC2 CC-series gap of 2026: how AI is governed inside the org.
4. Open Policy Agent (OPA) — control-as-code. Once your policy decisions are written down in plain English, you encode the ones that can be automated as Rego policies. OPA lets you express "production data may not flow to model X" or "no SSH access without an open ticket" in a single language and enforce it across services. The audit evidence is the Rego file plus the decision log.
5. CloudQuery — Sync Cloud Infrastructure to SQL for Security and Compliance — your control mapping evidence engine. CloudQuery pulls inventory and configuration state from AWS, Azure, GCP, Okta, GitHub, and dozens of other systems into a SQL database. Now your control "S3 buckets must be private" becomes a query you can re-run for the next audit. Auditors love SQL evidence; auto-generated screenshots they tolerate.
6. Presidio — Detect and Anonymize PII — the DLP layer between sensitive content and any AI workflow. Microsoft's open-source PII detection and anonymisation library. Pipe risk-assessment narratives, customer support transcripts, or evidence excerpts through Presidio before sending them to a cloud model. Reduces the surface area on which a vendor incident becomes your incident.
7. Wazuh — Open Source XDR & SIEM Security Platform — continuous monitoring (SOC 2 CC7) and incident detection in one open platform. The compliance team rarely operates Wazuh directly, but you need a SIEM whose audit evidence (alert dispositions, log retention, file-integrity records) you can export at audit time. Wazuh is the open-source option you can self-host so the evidence corpus stays under your control.
8. Immudb — Immutable Database with Cryptographic Verification — your tamper-evident audit log. Compliance regimes increasingly demand that audit logs be append-only and cryptographically verifiable. Immudb writes every record with a Merkle-tree proof; you can show an auditor exactly what was logged when and that nothing has been silently edited since.
9. Bernstein — Audit-Grade Orchestrator for CLI Agents — the wrapper that turns ad-hoc agent runs into auditable evidence. When the Compliance Auditor agent generates a draft gap analysis, Bernstein captures the prompt, the model, the inputs, the outputs, and a chain-of-custody trail. This is how you answer the auditor question "how do you know the AI's analysis here is reproducible?".
10. Guardrails AI — Validate LLM Outputs in Production — the output validation layer for any AI-generated artifact you'll attach to evidence. Schema-validates LLM responses, blocks hallucinated control IDs, and refuses outputs that don't conform to the structure your audit workflow requires. Most compliance hallucinations are caught here, not in human review.

How they fit together

  Policy docs / contracts / evidence ─► Claude Compliance Auditor
                                              │
                                              ▼
                       Regulatory-Checks Compliance Auditor (sibling)
                                              │
                                              ▼
      ┌───────────── Bernstein wrapper (chain-of-custody) ─────────────┐
      │                                                                 │
      │   ┌─ Agent Governance Toolkit ──► policy gap analysis           │
      │   ├─ OPA ───────────────────────► control-as-code               │
      │   ├─ CloudQuery ────────────────► control map + evidence (SQL)  │
      │   ├─ Presidio ──────────────────► PII redaction before any call │
      │   └─ Wazuh ─────────────────────► continuous monitoring evidence│
      │                                                                 │
      └───────────────────┬─────────────────────────────────────────────┘
                          ▼
              Guardrails AI (schema-validate every AI output)
                          │
                          ▼
                Immudb (immutable, cryptographically-verified log)
                          │
                          ▼
               Human-signed draft report → auditor walkthrough

Tradeoffs you'll hit

• One unified compliance suite vs this open stack. A managed GRC platform (Vanta, Drata, Secureframe) is faster to stand up and includes pre-built control mappings. This pack wins on data sovereignty (your evidence stays where you control it), avoids per-seat lock-in, and lets you encode controls that don't fit a vendor's playbook. Most enterprises end up running both: the managed platform for the standard controls and an open stack like this for the long tail.
• AI-assisted gap analysis vs auditor-led. A frontier model can read your policies and a framework's control catalog and surface plausible gaps in minutes. It will also fabricate plausible-sounding control IDs. Use the AI for the first pass; verify every cited control against the actual framework text; never let an AI-only analysis go to the auditor.
• Cloud frontier model vs local-only. Some of these picks (Presidio, OPA, Wazuh, Immudb, CloudQuery) are entirely local infrastructure with no LLM. Others (the Compliance Auditor agents, Bernstein, Guardrails AI) presume an LLM. For the LLM-touching steps, the default posture is: redact PII via Presidio, send to your enterprise-contracted model (zero-retention terms), capture the trace via Bernstein. Consumer chatbot tabs do not appear anywhere in this workflow.
• Immutable audit log vs ordinary database. Immudb is slower to write than Postgres and harder to operate. The reason to take that cost: when an auditor asks "prove this log was not edited after the fact", a regular database cannot answer the question. If your regime doesn't demand cryptographic verifiability, a normal append-only table is fine.

Common pitfalls

• Treating AI gap analysis as the gap analysis. The agent will surface 40 plausible gaps. Some are real, some are paraphrases of the same gap, some don't apply because of a control you haven't told it about. The output is a starting point for the compliance team to triage, not the deliverable.
• Sending raw policy text to a consumer chatbot. Internal policies frequently contain customer names, vendor terms, system topology, and IP. None of that belongs in a free chatbot tab. Presidio plus an enterprise-contracted model is the defensible default.
• Auto-generating control evidence with no human review. "CloudQuery says all S3 buckets are private" is evidence only if a human checked the query and the scope. Auto-run, auto-attached evidence with no human in the loop is exactly the audit finding you want to avoid.
• Conflating AI hallucinations with controls. When the model invents a SOC 2 control ID ("CC-9.8.4") that doesn't exist, downstream documents inherit the fiction. Guardrails AI plus a strict schema for control references catches this before it lands in the report.
• Mistaking this pack for a substitute for an auditor. External auditors exist because independent assurance is the whole point of an audit. None of these tools — agents included — change that. They make the team's preparation cheaper, faster, and better documented; they do not produce the attestation.
• Letting the audit log become an afterthought. Immudb has to be wired in before the agents start running, not retro-fitted at the end. If your chain-of-custody is "we exported the chat afterward", that is not chain-of-custody.