Accenture Cuts Clients' Bedrock Token Costs by 40–85%
The price per token is falling, but AI bills are rising: agents make many model calls where one used to be enough. Across four client projects, AWS and Accenture showed how to cut spending by 40–85%, for example by no longer paying again for the same long instruction in every request. A telecom operator reduced its costs from $12M to $3.8M a year (−68%). Start with cost tags, and enable the techniques one at a time while measuring quality.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Tokens keep getting cheaper, yet AI bills keep growing. Cheap tokens attract new workloads, and agents make several model calls where one used to be enough. Picture a support chatbot with a long system prompt: it goes to the model with every message, and you pay to process it again each time.
Multiply that by thousands of conversations, and consumption outpaces the drop in prices. This is the problem AWS and the consulting firm Accenture propose to tackle on Amazon Bedrock. They described six cost-saving techniques.
On four client projects (telecom, insurance, mining, and a large enterprise deployment), token spending fell by roughly 40–85% where it could be measured. If you pay for model APIs, this is a ready-made playbook. The industry itself already considers the topic important: the Linux Foundation has launched the Tokenomics Foundation to develop open standards for AI economics.
First, understand who spends what
Without tagging, the AWS bill shows AI as a single line item, with no breakdown by team, agent, or model. So the first step is cost tracking. For each application, you create a separate application inference profile and tag it: team, cost center, environment. Then you activate the tags as cost allocation tags in the AWS Billing and Cost Management console.
Three sources supply the data. Cost and Usage Reports (CUR 2.0) show how much money each workload and team spent. CloudWatch counts calls and latency. The Bedrock invocation log records the model and the number of input and output tokens in every request.
For product teams, Accenture suggests a metric called Cost per Business Transaction (CpBT): the full cost of one successful business outcome. It includes all model calls, retries, tool calls, and handoffs between agents. This exposes the cost of failed attempts, which the per-token price hides.
Six techniques in three phases
Phase 1: settings only. Prompt caching lets you avoid paying the full price for repeated parts of a request: system instructions, tool descriptions, reference documents. AWS promises a discount of up to 90% on cached input for supported models. Batch inference runs non-urgent tasks asynchronously and costs 50% less than the standard mode for a number of models. It suits overnight reports, document processing, and quality-evaluation runs.
Phase 2: architecture. Intelligent prompt routing sends simple requests to smaller models within the same family and promises up to 30% savings without loss of accuracy. Check the cost of the routing itself in the price list. The second technique is RAG optimization (where the model answers based on retrieved documents): hierarchical document chunking, metadata filters, and reranking, so that only the relevant pieces reach the model. Accenture adds two techniques of its own: each agent receives only its own fragment of the documents, and documents are compressed before being sent to the model.
Phase 3: infrastructure and training. A semantic cache returns a stored answer to a similar question, even if it is worded differently. On AWS, you build it on a vector store such as Amazon ElastiCache placed in front of the model. In an AWS test on 63,796 real chatbot queries with a 90% similarity threshold, the cache hit in 74.5% of cases, accuracy was 92.3%, and savings were 72.5%. Distillation transfers knowledge from a large model into a small one. According to AWS, such models in Bedrock are up to 500% faster and up to 75% cheaper than the original, while accuracy drops by less than 2% on tasks like RAG. It pays off on stable, high-volume tasks.
What the clients achieved
A large enterprise customer was sending every request through the most powerful model, with no caching or batching. Accenture enabled model-by-task routing, prompt caching, and batch processing at the same time. Monthly token spending fell by 85%. About 15% of tasks still required the top model.
A telecom operator was spending $12M a year: agents spawned requests without limits, and everything went to the most powerful model. After complexity-based routing, a semantic cache, and context compression, input tokens fell by 70%, and annual spending dropped to $3.8M (−68%). Quality was checked against the operator's existing internal tests.
An insurance company enabled prefix caching at the model level, and a mining company enabled a semantic cache in the application. Together they saved about 40% without changing models and without training.
The caveats matter. This is a vendor blog, the clients' names are hidden, and in two projects the techniques were introduced simultaneously, so the contribution of each is unknown. Results depend on the workload.
What to do in practice
Apply tags, find the most expensive workloads through CUR 2.0 and the invocation log, and enable caching for interactive traffic and batch for asynchronous traffic. Before making changes, record baseline cost and quality, and verify them on a held-out sample of real traffic, not synthetic data. Change one technique at a time so you can roll back. In regulated industries, document test results and the rollback plan. With routing and semantic caching, keep Guardrails on every path and set a cache lifetime so users don't receive stale answers.
- Nov 19, 2026Two pods share one A10G GPU on Amazon EKS: NVIDIA Run:ai gave each 8 GiB
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026AWS AgentCore Agents Can Now Reach a Private Keycloak Behind a Corporate CA
Key Questions Answered
What are the key details on First, understand who spends what?
Without tagging, the AWS bill shows AI as a single line item, with no breakdown by team, agent, or model. So the first step is cost tracking. For each application, you create a separate application inference profile and tag it: team, cost center, environment. Then you activate the tags as cost allocation tags in the AWS Billing and Cost Management console. Three sources supply the data. Cost and Usage Reports (CUR 2.0) show how much money each workload and team spent. CloudWatch counts calls and latency.
What are the key details on Six techniques in three phases?
Phase 1: settings only. Prompt caching lets you avoid paying the full price for repeated parts of a request: system instructions, tool descriptions, reference documents. AWS promises a discount of up to 90% on cached input for supported models. Batch inference runs non-urgent tasks asynchronously and costs 50% less than the standard mode for a number of models. It suits overnight reports, document processing, and quality-evaluation runs. Phase 2: architecture.
What the clients achieved?
A large enterprise customer was sending every request through the most powerful model, with no caching or batching. Accenture enabled model-by-task routing, prompt caching, and batch processing at the same time. Monthly token spending fell by 85%. About 15% of tasks still required the top model. A telecom operator was spending $12M a year: agents spawned requests without limits, and everything went to the most powerful model.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.