How we cut AI costs by 80%
A sponsored experiment by Port shows that pre-integrating enterprise data into a unified 'context lake' reduces AI agent token costs by up to 80% compared to connecting agents directly to multiple MCP servers. The test ran 12,000 queries across four conditions and three Claude models (Haiku, Sonnet, Opus). Key findings: a context lake alone cuts costs ~58%, and adding a skill file (a routing table mapping query types to catalog fields) brings savings to ~80%. Counterintuitively, adding a skill file to raw MCP access made costs 13–24% worse, as agents followed it as a checklist rather than reasoning efficiently. The efficiency comes from pre-joined data (services already linked to their team, repo, PagerDuty, and Jira) and pre-computed aggregations, shifting relational reasoning from inference time to ingestion time. The post argues platform engineering teams should own context management as a budgeted resource.
