← All engineering notes

Enterprise AI

Review Bedrock inference profiles for data residency

Check where Bedrock can process an internal AI request when a model uses a geographic or global inference profile.

An AWS Region in an application configuration does not by itself settle where Bedrock processes inference. A Bedrock inference profile can route a request across Regions. The profile type, its current destination set, and the controls around it are the facts an internal AI team must review.

Amazon Bedrock describes geographic and global cross-Region inference separately. A geographic profile routes within its named geography, such as EU or APAC. A global profile can route across supported AWS commercial Regions. This is an engineering control review, not a legal conclusion about residency or sovereignty.

Start with the profile, not the source Region

Capture the exact model or inference-profile ID that each application calls. Then record the profile type and the Regions AWS currently lists for it. A source workload in Sydney can still make a cross-Region inference request if its selected profile allows it.

Review question Evidence to retain Failure if skipped
Is the ID a regional model or an inference profile? Deployment configuration and request trace The team reviews the wrong routing behavior
Is the profile geographic or global? AWS profile documentation and model configuration A global route violates an internal processing restriction
Which Regions are in the profile today? Dated export or checked AWS documentation A profile change is missed during a model update
Can organization policies allow every destination? IAM and SCP review A request fails only after rollout
Can operations identify the actual inference Region? CloudTrail query and retention plan An incident cannot establish the processing path

AWS states that geographic profiles select a commercial Region within the specified geography, while global profiles select from supported commercial Regions. It also notes that profiles can route to Regions that an account has not manually enabled. That makes the profile’s allowed set more useful than a console screenshot of enabled Regions.

Verify the route from a request record

Run a non-sensitive test request through each profile that an application intends to use. Keep the timestamp, source Region, profile ID, model ID and returned request identifier. Then query CloudTrail in the source Region for the matching Bedrock event. AWS says cross-Region inference records the destination in additionalEventData.inferenceRegion.

CloudTrail events have a structured record format and can contain request and response details according to the service and event type; see the AWS event record reference. Design the evidence process around identifiers and configuration, not around storing employee prompts. The trace should establish route evidence without becoming a second content store.

Repeat the test after a profile, model or organization-policy change. A passing test proves only that one request used an allowed destination at that time. It does not prove every future request, background job or new profile will do so.

Make fallback obey the same boundary

Availability work often introduces the residency mistake. A team starts with an approved model, then adds a fallback model or a global profile after a throughput incident. The fallback becomes the least-reviewed route.

Create a routing matrix before enabling fallback:

Logical use case Primary profile Fallback Permitted geography Action if neither route qualifies
Employee search summary Hypothetical apac profile None until reviewed APAC Refuse and queue a manual review
Public knowledge assistant Hypothetical regional model Reviewed global profile Global Retry within the request deadline

The names in this table are examples, not Bedrock configuration advice. The important choice is explicit behavior when no destination is compliant. “Use any available model” is a routing instruction that can conflict with a data policy.

Alectura’s routing policy can limit each attempted connection by model, region, data terms and connection, and it does not perform mid-stream fallback. Review the platform routing description against the Bedrock profile and AWS account controls; the gateway cannot convert a provider profile into a legal guarantee.

Test the assumptions that fail in production

Test a permitted profile, a global profile that your policy should reject, a fallback path, and an organization policy that denies one listed destination. Check both the application result and the CloudTrail evidence. Include background workloads, batch paths and incident-response scripts in the inventory.

Geographic routing can support a bounded engineering rule. It does not replace a review of retention, provider terms, connected data sources, or local requirements. Use the Alectura technical coverage to test request-level policy enforcement with a non-production application, then keep ownership of the wider data-flow decision with your security and legal teams.

Sources checked 10 November 2026. Provider capabilities and terms can change; check the linked documentation for your configuration.

Keep reading

Design request identifiers for LLM retries and billing ↗Build an evaluation set for sensitive-data protection ↗Handle policy refusals and interrupted LLM streams ↗