Companion Article · Coming June 2026

The AI Tax.

What an agent-driven KPI stack actually costs - and the five patterns that cut it 70% without losing capability.

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

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. The 90% cost reduction most teams skip, and why it's tailor-made 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. 80% of LLM calls eliminated, same coverage.
  • 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 reducing its monthly bill by 70-80% step by step, with the actual prompts and Apps Script changes.
  • The compliance angle. How cost discipline ties to SOC 2, ISO 27001, and DORA cost-control requirements for regulated industries.
Want to read it the day it ships?

It will publish in mid-to-late June 2026. 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.