AI Effectiveness
Home Thesis Framework Journal

Series

Context Engineering The Sparsity Revolution The Art of Forgetting The Decision Effectiveness Series The Compounders

By Topic

Context Engineering Unstructured Data & RAG Agents & Emergence Data Governance
Labs About Subscribe
LinkedIn ↗ RSS Feed
← Back to Labs
Organization Live

The Compounding Test

Five questions. Live verdict. The diagnostic operationalizes The Berkshire Test for AI: one question about whether the AI is at the bottleneck of the workflow (Goldratt's Theory of Constraints), four about whether the four compounding conditions hold (Munger, Iansiti, Brynjolfsson, Perez). Your answers stay in the URL, so a verdict is shareable; the scoring is deterministic and runs entirely in your browser.

Location (Goldratt) — Is your AI applied at the bottleneck of the workflow?
1. Proprietary data origin — is the data generated by your operation, not bought or shared?
2. Self-labeling workflow — does the work itself produce labels (outcomes that grade the model's predictions)?
3. Decreasing marginal cost — is each model iteration cheaper than the last?
4. Defensible asymmetry — would a competitor need years (not months) to catch up?

One-shot win

Right place, conditions weak. A real gain at the bottleneck, but the flywheel won't start.

location 0.50 · compounding 0.25

Per-condition breakdown
Proprietary data origin
0.25
Currier, NFX (2022) — data network effects criterion #5
Self-labeling workflow
0.25
Iansiti & Lakhani, HBR (2020) — AI factory
Decreasing marginal cost
0.25
Brynjolfsson, Rock & Syverson, NBER (2018) — productivity J-curve
Defensible asymmetry
0.25
Perez (2002) — deployment-phase advantage

⚠ A one-shot win at the bottleneck is a real gain. Without proprietary data and self-labeling, the model plateaus quickly; what looks like a flywheel is a data scale effect. Casado & Lauten — The Empty Promise of Data Moats (a16z, 2019)

Your initiative most resembles

Mastercard

The network of labeled outcomes — every chargeback labels a transaction.

Read the full portrait →

How to read this

The verdict tile names one of four quadrants. Compounder is the only quadrant where capital deployed today compounds tomorrow. The other three are real failure modes, each with a different shape and a cited source for why it fails.

One-shot win: a real gain at the bottleneck without the compounding mechanism — competitors close the gap in roughly a year.

Compounding the wrong thing: a self-improving model at a non-binding step. The model gets better on its own metrics, but throughput at the bottleneck does not move.

Roman Candle: wrong location, no compounding mechanism. The Buffett-2007 failure mode: a moat that has to be continuously rebuilt will eventually be no moat at all.

For the full argument behind each axis, the four conditions, and four portrait companies that pass the test at scale (Progressive, Deere, Mastercard, Mayo Clinic), read the anchor essay: The Berkshire Test for AI .

Get the weekly briefing