From the source
- Accenture and AWS are laying out six levers for cutting token spend on
Amazon Bedrock, after finding that enterprise AI bills often climb even as per-token prices fall.
- Accenture applied the levers in four client engagements across telecom, insurance, mining and a large enterprise deployment, and says token spend fell by about 40–85 percent where savings were measured, though results vary by workload.
- The first step is attribution: cost allocation tags applied through Bedrock application inference profiles, one per application or workload, tagged by team, cost center and environment, then activated in the AWS Billing and Cost Management console.
- AWS Cost and Usage Reports, CloudWatch and Bedrock model invocation logging supply the telemetry, while Accenture's Cost per Business Transaction framework rolls invocations, retries, tool calls and agent handoffs into one fully loaded cost per successful business output.
- Phase one needs no infrastructure change: prompt caching reuses repeated prompt prefixes and is designed to cut cached input token costs by up to 90 percent for supported models, and batch inference runs non-urgent work asynchronously at 50 percent lower cost than on-demand pricing for select models.
- Phase two covers architectural changes, including intelligent prompt routing across models in the same family, which AWS says can reduce costs by up to 30 percent without compromising accuracy, and RAG optimization to keep unnecessary context tokens out of the model.






