Article · Liz Eversoll and Joe Shepherd

AI's real cost is not compute. It is context.

AI does not get expensive because you use too much of it. It gets expensive when it runs without understanding the work.

Inside this article

  • The wrong metricWhy token consumption is a vanity measure.
  • Why AI rediscoversKnowledge that exists but cannot be applied.
  • What changesReasoning moves into the knowledge layer.
  • The first moveOne costly workflow, encoded and measured.

Agents infer, approximate, and reinterpret the same logic on every run. That is what drives both inconsistency and cost.

This is a summary. Read the full article

The wrong metric

Uber's CTO told The Information the company burned through its entire 2026 AI coding budget in four months. Goldman Sachs projects agentic AI will drive a 24-fold increase in token consumption by 2030. Meanwhile engineers at large technology companies are encouraged, sometimes through internal leaderboards, to maximize token consumption as a proxy for adoption. Cognizant's CEO called that what it is: a vanity metric.

Token consumption is Goodhart's Law again. When a measure becomes a target it stops being a good measure. More prompts and more agents do not produce better decisions.

Why AI keeps rediscovering what you already know

Most enterprise knowledge already exists — policies, frameworks, data models, domain definitions. It is not structured in a way AI can apply. Documents describe intent but cannot enforce it. Data is structured, but the constraints governing how it behaves are not. Prompts try to fill the gap and add inconsistency every time.

So agents retrieve information without applying it systematically. They infer, approximate, and reinterpret the same logic on every run. That is what drives both inconsistency and cost.

What changes

SIGN™ moves the reasoning out of the model and into the knowledge layer. It defines what must happen, what cannot happen, and how conclusions are derived, so agents operate against governed logic instead of inference. Prompts get shorter. Workflows become reusable. Outputs stop varying run to run.

Today

Inference on every run

  • Documents describe intent but cannot enforce it.
  • Data is structured; the constraints on how it behaves are not.
  • Prompts try to fill the gap and add inconsistency every time.

ResultThe same logic, reinterpreted on every run.

With SIGN™

Governed logic

  • What must happen.
  • What cannot happen.
  • How conclusions are derived.

ResultShorter prompts, reusable workflows, outputs that stop varying run to run.

The first move

For a CTO the first move is small and measurable: take one high-cost agentic workflow, encode the knowledge that governs it, and watch the token count.

Read the full article

This page is a summary. The full article has the complete argument and every example.

Related

Explainer

What is SIGN™

Prose an agent has to interpret, next to the same rule expressed as SIGN™.

Explainer

Why this is not retrieval

What happens to a policy answer when chunks come back out of order, and what deterministic retrieval does instead.