Comparison

    Bedrock Model Invocation Logging vs Vertex AI Request-Response Logging: Do Provider Logs Count as Audit Evidence?

    Provider-native logs are supporting records, not audit evidence on their own. Amazon Bedrock model invocation logging and Vertex AI request-response logging record what the provider observed, under per-service configuration. They do not show which policy applied to an interaction or whether a control was enforced. Evidence that holds across providers needs a record linked to the enforcement decision, in a consistent format, with protected storage and defined retention.

    Evidence requirements that hold across providers

    • A record of the request and response with metadata, for every in-scope interaction
    • The policy that applied to that specific interaction
    • The enforcement outcome: allow, redact, block or escalate
    • One consistent schema across all models and providers
    • Protected storage with restricted access, defined retention and sufficient capacity
    • Retrieval on request by an auditor or authority, from logs the enterprise controls

    A log is not yet evidence

    Amazon Bedrock model invocation logging and Vertex AI request-response logging each record request and response activity for models run on that provider. Both are configured per service. They differ in what they capture, where records are stored, how they are scoped or sampled, and how long they are kept. Those differences, not the existence of logging, determine how far a team can rely on them.

    NIST's Cybersecurity Framework Profile for AI describes AI audits as demonstrating compliance with legal, regulatory and contractual requirements, and points to logging that shows requests, responses, metadata and metrics. A provider log maps well to that description. An auditor usually asks for more: whether the record is complete, who could alter it, how long it is retained, and whether it connects to a decision about what the AI was permitted to do.

    The useful test is therefore not whether a provider produces logs. It is whether each log, as configured in your environment, holds up against those questions.

    Attributes to test in each provider's logging

    Apply the same attributes to both services, using your own configuration rather than defaults. Record the answer for each provider side by side. Differences in the answers are the audit risk in a multi-provider estate.

    • Capture scope: Which models, endpoints and interaction types are covered. Whether sampling or payload size limits apply. Whether prompts, responses and metadata are all retained.
    • Configuration and enablement: Logging is a per-service setting that must be switched on and kept on. Ask who can change it, whether changes are themselves recorded, and what exists for interactions before enablement.
    • Data destinations: Where records land, such as Amazon S3, CloudWatch Logs or BigQuery, in which account or project, and under whose administrative control.
    • Retention: Retention is governed by the destination. Match it to the period your applicable framework requires, and plan storage capacity so logging capability is not lost.
    • Access control and integrity: Who can read, modify or delete records. Whether log access is separated from the teams generating the logs. Whether tampering can be detected.
    • Gaps: Policy applied, enforcement outcome, and a consistent schema across providers. A provider log is not designed to answer these.

    Where provider logs support and where they fall short

    Provider logs do real work. They let a team reconstruct what a model received and returned, investigate an incident, and cross-check other records. For this purpose their proximity to the model is an advantage. They should be kept and protected as one input to an evidence strategy.

    They fall short at the points an auditor tests hardest. NIST SP 800-144 says audit mechanisms should help verify policy enforcement, and notes that cloud consumers can depend heavily on the provider for event data. A log of what the provider saw does not by itself demonstrate that a policy was enforced. If a request contained regulated data and the response was clean, the log cannot say whether a rule redacted it, whether none applied, or whether nothing was checked.

    In the EU AI Act's consolidated text, logging capabilities for high-risk AI systems must enable recording of events relevant to traceability, and providers must give a competent authority access to automatically generated logs to the extent those logs are under their control. As of November 2026, that makes control over the logs a practical question. Whether a given workload falls within those provisions, and in which role, depends on the facts of the deployment.

    Running two providers compounds the problem. Each has its own scope, format and retention settings, so a single audit question can require two queries against two schemas, with different coverage.

    How runtime enforcement produces the missing evidence

    The first three requirements above cannot be reconstructed reliably after the fact, because the decision happens before the response reaches the user. Evidence of enforcement has to be produced by the component that enforces.

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Each interaction or tool call is allowed, redacted, blocked, rate limited or escalated for human approval, and the decision is enforced at runtime, not reviewed after the fact. The audit-ready evidence is generated by that enforcement and mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST. Because one policy plane covers more than 100 models in production, the record format does not change when the provider does.

    Trussed deploys self-managed in the customer cloud, with configurable retention and data residency, so the enterprise holds the records. This does not replace provider-native logs. They remain a useful supporting record of what each provider observed, and an enforcement-linked record adds what they cannot: the policy applied and the outcome.

    From provider log to audit evidence

    Provider log

    Request and response as the provider observed them

    Policy applied

    Which rule governed this interaction

    Enforcement outcome

    Allow, redact, block or escalate

    Evidence record

    Consistent, protected and retained across models and providers

    Test your evidence against what auditors ask

    See how enforcement in the execution path produces audit-ready evidence that complements provider-native logs across models.

    Request a Demo