Companion Article · In production

What agent-driven measurement actually costs

What an agent-driven KPI stack can cost - and five patterns that cut it by an estimated 70-80%.

Unoptimized
$200-500
per month, for a 10-KPI portfolio at default agent settings
Optimized
$40-100
same capability, with model tiering, caching, batched investigation, statistical filtering

Both ranges are the author's illustrative estimates for a 10-KPI portfolio, not measured results. Actual cost depends on the model, the prompt size and how often each agent runs.

What the full article will cover

  • Model tiering economics. When to use Haiku (routine), Sonnet (medium complexity), Opus (hard cases). The cost-per-call differential and how to route automatically.
  • Prompt caching. Cached input tokens cost 10% of the normal input price for most Claude models, a saving most teams skip, and a natural fit for the Investigation Agent's recurring context.
  • Batched investigation across the portfolio. One agent reading ten metrics vs ten agents reading one each. The math and the architectural pattern.
  • The statistical filter pattern. Cheap statistical anomaly detection as a pre-filter; LLM analysis only on suspicious rows. The share of calls this removes is measured in the worked rewrite.
  • The cost-per-decision metric. Why measuring API spend per month is the wrong frame, and what to measure instead.
  • Worked rewrite. Taking the Investigation Agent from the Ship KPIs framework and targeting a 70-80% lower monthly bill, step by step, with the actual prompts and Apps Script changes.
  • The compliance angle. How spend caps and API usage logging support audit-logging and vendor-oversight controls in regulated environments.

Sources

  • Anthropic, Prompt caching, Claude Platform Docs. Cache read tokens are 0.1 times the base input token price, with per-model exceptions.
  • Anthropic, Batch processing, Claude Platform Docs. The Message Batches API cuts cost by 50%; most batches finish in less than one hour.
  • Author's estimates: the monthly cost ranges ($200-500 unoptimized, $40-100 optimized) and the 70-80% reduction are illustrative, not measured.
Want to read it the day it ships?

The full article is in production. In the meantime, the framework it builds on - Ship KPIs Like Features - is the prerequisite read.

Read the Framework Article →

This page is a placeholder so the references in the article and LinkedIn series resolve to a real URL. The full article is in production. Bookmark this page or follow May on LinkedIn for the publication notice.