Research: How we cut AI token costs by 80%

A study conducted by Zohar Einy reveals that managing AI token costs can be significantly reduced by utilizing a "context lake" system, which creates a unified knowledge graph that integrates data from various tools like GitHub, Jira, and PagerDuty. This system reduces the need for repetitive and costly data queries by pre-integrating and organizing data, thus lowering token consumption by approximately 80% compared to traditional methods that rely on multiple MCP server calls. The experiment tested different configurations, including the use of skill files, and found that while straightforward skill files increased costs in some setups, they effectively reduced costs when combined with a context lake. The research suggests that platform engineering teams should manage context resources to optimize costs and efficiency, moving work from inference time to ingestion time, thus allowing organizations to handle complex queries more economically.