AWS has published a deployable data-lake design that brings Amazon Bedrock invocation logs from linked accounts into one queryable view. It lets teams inspect individual prompts and responses, track token use by model and account, and join usage to cost data. That is useful visibility, with a rather important privacy footnote: the same pipeline centralises the conversation text.
AWS AI Watch analysis
What happened
The design routes Bedrock invocation logs through CloudWatch and a central account, then processes them with Lambda and Firehose into Parquet files in S3. Teams can query the resulting data in Athena, including the prompt, model response, input and output token counts, stop reason, account, timestamp and IAM role for each invocation.
AWS says teams can join the processed data to Cost and Usage Report 2.0 using the IAM ARN to attribute spend. Its sample also includes CloudFormation templates, code and SQL queries. Logging must be enabled separately in each source account; deploying the stacks alone does not start it.
Why it matters
A single view can help operators investigate a specific call, compare token use across teams and models, find cut-off responses, and connect usage to cost. That is a more actionable picture than an aggregate bill, especially for organisations running Bedrock across multiple accounts.
But this is not just a dashboard for tidy charts. The data lake can contain full prompts and completions, which AWS notes may include personally identifiable information. Its design uses KMS encryption and IAM access controls, so decisions about permissions, retention and who can query the central store matter as much as the SQL.
Our read
This is a practical observability blueprint, not a magic switch: teams need to deploy the account infrastructure, enable logging and build the Athena view. The payoff is clearer evidence about what models are doing and what they cost. The trade-off is that sensitive conversation data becomes easier to collect and search centrally. Useful plumbing, then, with governance very much part of the plumbing.
What to watch
- Whether AWS’s sample solution is maintained as Bedrock logging and billing formats evolve.
- How teams restrict access to prompts and responses, and how long they retain them.
- Whether organisations use per-invocation cost data to change model choices or prompt design.
Discussion spark: Would you centralise full prompts and model responses to improve cost and usage oversight, or keep that data spread across accounts to limit exposure?
Sources and evidence
- Amazon Bedrock tokenomics data lake (10 November 2026, 16:56 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)