What the checklist covers:
30
Evaluation criteria for audit logging across LLM, agent, and MCP traffic
5
Control areas, from event coverage to access controls and compliance mapping
17
Must-have controls, each flagged so you know what to fix before an audit
4
Regimes mapped: SOC 2, HIPAA §164.312(b), GDPR, and export-control recordkeeping
Inside the checklist
Download the checklist to get:
A working framework your platform, security, and compliance teams can use to evaluate an AI gateway or audit your own AI stack before the auditors arrive.
Download the Checklist✓
The full 30-point AI audit logging framework
Every criterion grouped by control area, with the question to ask and a must-have or should-have priority.
✓
A three-step rollout plan
Define which events to capture, set retention by compliance regime, and export securely to your SIEM or warehouse.
✓
A compliance evidence map
How audit logs support SOC 2, HIPAA audit controls, GDPR records of processing, and export-control recordkeeping.
✓
A vendor evidence-pack request list
The sample logs, RBAC settings, export configs, and retention settings to ask any AI gateway vendor for during review.
Why it matters
Every AI Request Is Now Audit Evidence
Enterprise AI applications sit between regulated data and external model providers. Every prompt, response, tool call, and policy decision that passes through them is potential evidence for SOC 2, HIPAA, GDPR, ITAR, and internal investigations.
Traditional application logs weren't built for this. They record that an endpoint was hit, not which user's data was sent to which model, which MCP tool an agent called with what parameters, or which guardrail blocked the response. Agents make it harder: a single request can fan out into a dozen non-deterministic steps, and if those steps aren't tied together by one trace, nobody can reconstruct what happened.
One request, many hops, one reconstructable record. The checklist defines what that record has to contain and where it should live.
Teams that try to retrofit audit logging onto AI usually hit the same gaps. Logs show a shared API key instead of a user. Payloads are either stored in full, exposing PII to anyone who can read the logs, or not stored at all. Agent steps and tool calls live in separate systems with no shared correlation ID. And the logs sit in a vendor's infrastructure, outside the retention and access policies your auditors expect.
What good AI audit logging looks like
Every request carries the authenticated user, the calling application, and tenant and environment tags. Prompts, responses, tool calls, and agent steps are captured at a level set by data class, with PII redacted at write time. Guardrail hits, budget blocks, and authorization denials are logged with the rule that fired, and every model, prompt, and policy version is stamped so incidents can be reconstructed.
The logs live in object storage you own, with retention and legal holds mapped to your compliance regimes. They stream through OpenTelemetry to your SIEM and data warehouse in near-real-time, and access to them is controlled by role, including scoped, time-bounded access for external auditors.
Logging is centralized at the AI gateway
Instrumenting every application and agent separately leads to inconsistent coverage. Routing LLM, MCP, and agent traffic through an AI gateway gives you one enforcement point where identity, redaction, policy decisions, and trace context are applied the same way for every request.
The AI Audit Logging Checklist turns that standard into 30 criteria you can use to evaluate a gateway or audit your current stack. Download it to find the gaps before your auditors do.
Preview
The five control areas in the AI audit logging checklist
A sample of the criteria in each area. The full checklist includes all 30, with priorities and evaluation questions.
Area 1 · 7 checks
Event coverage
Tool calls and MCP invocations are audited with parameters and results
Guardrail hits, budget blocks, and authorization denials log the rule that fired
+5 more checks
Area 2 · 7 checks
Identity and attribution
The authenticated user or service account is on every request log
Correlation IDs propagate across app, gateway, agent, and provider
+5 more checks
Area 3 · 5 checks
Storage, retention, and ownership
Audit logs can be stored entirely in customer-owned object storage
Retention is configurable per team, app, or data class
+3 more checks
Area 4 · 5 checks
Export and SIEM integration
Traces, metrics, and logs export to OTEL-compatible systems
Logs flow to Splunk, Datadog, Sentinel, or other SIEMs
+3 more checks
Area 5 · 6 checks
Access controls and compliance
Log access is restricted by role: security, ops, audit, developer
PII is redacted from logs at write time
+4 more checks
How to use · 3 steps
Rolling it out
Define events: auth, prompts, responses, tool calls, agent steps, policy decisions, errors
Set retention by regime (SOC 2, HIPAA, GDPR) and investigation needs
Export securely to SIEM or warehouse, with access controls and tamper-evidence
4.7
★★★★★
Gartner Peer Insights
Rated 4.7 on Gartner Peer Insights
Platform teams run TrueFoundry's AI Gateway for LLM, MCP, and agent traffic in production, with OTEL-compliant request logs, traces, and metrics, RBAC-protected log access, and a logging architecture that keeps audit data inside their own cloud.
96%
of reviewers likely to recommend
4.9
ease of deployment and admin
3–4 ms
added gateway latency
