Exam-style
An Australian insurer runs a customer service assistant on Amazon Bedrock, deployed in the ap-southeast-2 Region because a contract requires that customer data stay in Australia. Legal asks the team to confirm in writing that the requirement is met. What should the team say?
Reveal the answer
B. The Region choice is necessary but not sufficient. A cross-Region inference profile can serve a request from elsewhere in the geography, CloudWatch Logs groups never expire by default, and copies taken for evaluation or support can land in a bucket in another Region
Choosing the Region is the primary residency control and it settles where the model runs and where the buckets sit, which is why the first option is tempting: for a workload with one copy of the data and nothing configured to move it, the Region really would be the whole answer. A generative AI feature is not that workload. It leaves a fresh copy at each stage, and three of those copies escape the Region setting. An Amazon Bedrock cross-Region inference profile spreads load across Regions within a geography, so a request submitted to the Sydney client can be served from another Asia Pacific Region; somebody enabled it during a busy week to get past throttling, and it needs to be checked or left off. CloudWatch Logs groups default to Never expire, so an application’s prompt and completion logging quietly becomes the longest-lived copy of exactly the text the contract is about, and it needs an explicit retention setting. And a copy taken for evaluation, for fine-tuning, or to attach to a support case lands wherever the person who took it put it, which is another Region unless the buckets themselves are constrained. The third option overshoots: prompts sent to a Bedrock model are not used to train it and are processed in the Region called, so the service is usable under the contract. Encryption and an AWS Artifact report are both worth having and neither answers where the bytes are: KMS controls who can read a copy, not which country holds it, and Artifact supplies AWS’s own audit evidence rather than proof about this workload.
Q. The assistant is deployed in ap-southeast-2 because customer data must stay in Australia. Legal wants that confirmed. Is the Region enough?
A. No. The Region choice is necessary and not sufficient. Check the inference profile, the log retention, and the copies people take.
Why? Residency is about where each copy of the data physically sits, and a generative AI feature makes a copy at every stage of its data lifecycles: the prompt, the retrieved passage, the completion, and the log line that records the exchange. The Region setting covers the first three and misses three ways out. An Amazon Bedrock cross-Region inference profile can serve a request from any Region in its geography, so either confirm which profile is in use or invoke the model directly in the Region. Application logging into Amazon CloudWatch Logs defaults to Never expire, which makes the log group the longest-lived copy in the system; set a retention period on it deliberately. Copies taken for evaluation, for fine-tuning or for a support case need buckets that are already in the right Region, with S3 lifecycle rules to move them into S3 Glacier storage classes and then delete them on a schedule nobody has to remember. An AWS Config rule watching the log groups and the buckets is how the team finds out when one of those assumptions stops being true rather than discovering it during the next review.