AI
Context Lakes Cut AI Agent Costs by 80%, Study Shows
Pre-integrated data and unified knowledge layers dramatically reduce token consumption in enterprise AI workflows, making expensive models more practical for routine tasks.
The New Stack
Enterprise AI costs spiral not just because models grow pricier, but because agents waste tokens on redundant data-fetching. A new study testing four different architectures found that pre-structuring organizational data into a unified layer—what Port calls a "context lake"—slashes costs by 80% compared to agents juggling multiple disconnected tools.
The Cost Problem#
Agents tackling common enterprise questions repeat the same queries thousands of times monthly. To answer "Is this service outage linked to a recent deploy?", an agent today calls GitHub for ownership, Jira for tickets, and PagerDuty for incidents—three separate hops, each one reinvented from scratch. Context accumulation drives costs, according to Gartner, as growing context windows compound with every agent step.
The Experiment#
Researchers tested four setups across three Claude models:
- Claude + MCP servers (baseline) – Direct access to GitHub, Jira, PagerDuty; agents figure out which tool to call and reconcile mismatched formats themselves
- Claude + MCPs with skill file – Same tools, plus a markdown guide on lookup strategy
- Context lake – Unified catalog where services, teams, incidents, and deployments are pre-linked
- Context lake + skill file – Catalog with optimized routing instructions
Running 12,000 queries across all conditions showed dramatic differences:
- Haiku: $0.087 → $0.018 (79% savings)
- Sonnet: $0.333 → $0.059 (82% savings)
- Opus: $1.761 → $0.354 (80% savings)
Surprisingly, adding a skill file to raw MCPs made costs worse (13–24% higher), as agents treated instructions as checklists rather than reasoning aids. The skill file only helped when data was already well-structured.
Why Context Lakes Win#
Two factors drive efficiency:
Pre-joined data. A service entity already knows its team, repo, PagerDuty service, and Jira project. The agent reads one node; the full picture follows. No mid-flight reasoning about how things connect.
Data shortcuts. Mirror fields like and aggregates like are computed once at ingestion, not repeatedly at inference time. The agent reads a value instead of counting.
