What changed
Accenture has shared a sequenced approach to optimizing AI economics on Amazon Bedrock, detailing six levers that can reduce token spend by up to 85% in some cases. These strategies are based on real-world client engagements across telecom, insurance, mining, and enterprise sectors. The methods include configuration changes, architectural adjustments, and infrastructure investments, all aimed at aligning AI costs with business outcomes.
What this means in practice
Accenture applied these optimization levers in four client engagements, with results varying based on each workload. The approach starts with configuration-only changes, such as prompt caching and batch inference, which can reduce input token costs by up to 90% and lower pricing for non-urgent workloads by 50%, respectively. These methods require no new infrastructure but need specific prompt formatting or batch job submission.
Architectural changes, like intelligent prompt routing, direct requests to models based on task complexity and business value, potentially reducing costs by up to 30%. Accenture also used retrieval optimization techniques to limit unnecessary context tokens, improving cost efficiency.
In more advanced phases, semantic caching and model distillation were applied, with the latter offering up to 75% cost savings and 500% speed improvements for certain workloads.
Limitations and availability
The optimization methods are available natively on Amazon Bedrock, with some requiring specific configurations or additional services like Amazon ElastiCache for semantic caching. Results are specific to each client’s workload and may vary. Client names are withheld for confidentiality. The Linux Foundation’s Tokenomics Foundation is noted as a developing standard for AI cost economics, but the methods described are based on Accenture’s implementation rather than industry-wide benchmarks.
