Agent Spend Control — What Did We Get For It?

Agent Spend Control · What Did We Get For It?

We could always see what the agents spent. Never what we got for it.

MODELING LIVE

UNGOVERNED SPEND / MONTH ($)

INFERENCE SHARE (%)

PREMIUM-DEFAULT SHARE (%)

DEFLECTED THIS MONTH · GOVERNED VS UNGOVERNED

$1.27Mof $3.0M ungoverned

Governed CVU $0.18 ↓43% vs $0.32 ungoverned · $15.2M projected annual · every dollar tied to a validated result.

Ungoverned Governed Deflected

POLICY CONTROLS · TOGGLE TO SEE WHAT EACH IS WORTH

WHAT'S HAPPENING—

GOVERNED CVU

$0.18

↓43% per validated unit

DEFLECTED · MONTH

$1.27M

declined · downgraded · routed

PROJECTED · ANNUAL

$15.2M

at current policy mix

VALIDATED UNITS · MO

9.6M

cleared the quality bar

CONTROL OVERHEAD

0.8%

of governed spend · sub-linear

LIVE DECISION STREAM click a row

PROVIDER RANKING $/validated unit

WHAT COUNTS AS A RESULT the quality bar

10.0M

ATTEMPTS

9.6M

VALIDATED

1.4%

ESCAPED

A validated unit clears the acceptance bar — in a controlled enrichment test the cheaper path verified 96.4% vs 91% manual. CVU counts cost over validated units only.

AI · POLICY3 standard-criticality agent classes default to premium scoring with no validated-quality lift. Apply the mini policy — −$0.7M/mo.

AI · ROUTINGFor enrichment, Provider B clears the bar at $0.04 vs $0.07. Shift 60% of volume — CVU −18%.

ASSUMPTIONS & SENSITIVITY — the model, shown

Ungoverned spend / month	$3.0M (McKinsey: production agents = $5–50M/yr)
Premium → mini price delta	~20–37× (frontier ≈ $15–30/M out vs mini ≈ $0.40/M, 2026)
Mini clears the quality bar	96.4% (controlled enrichment test vs 91% manual)
Downgradeable premium share	35% of inference (adjustable)
Retry waste	12% of inference spend (runaway loops)

Sensitivity: ±20% on premium-default share and price delta moves monthly deflection to roughly $1.0M – $1.5M. The thesis holds across the range — the question is the size of the prize, not whether it exists.