Chapter One

Same AI, Different Outcome

The cost of AI fell 300x in two years. The cost of getting it wrong stayed exactly the same. The gap between the two is where firms gain an advantage or lose money.

The cost of intelligence fell 300x. The cost of being wrong did not.

AI is now available to every firm at near-zero cost. AI costs have fallen close to 300x in two years. When the technology is identical, it stops being an advantage.

300×

Fall in AI cost in two years. The same intelligence, at the same price, to every competitor.

The question every leader is asking: what do we have that the firm next door does not?

The answer is your firm's judgement: The values, the policies, the goals, the relationships. It lives in your people and your culture, not in any model's training data.

Two problems follow. Most of it was never written down in a form a machine can use. And even when the machine uses it correctly, you have no way to prove that it did. Unwritten judgment and unproven output: that combination is what keeps AI investments from returning anything. This book sets out the discipline that fixes both.

Figure 1.1

Three firms, one identical model. The gap is the operating line.

FIRM A

APPROVAL PATHS

EXCEPTIONS

FIRM DATA

WRITTEN RULES

THE MODEL

FIRM B

NOT AGREED

NOT WRITTEN DOWN

FIRM DATA

WRITTEN RULES

THE MODEL

FIRM C

NOT AGREED

NOT WRITTEN DOWN

SCATTERED

WRITTEN RULES

THE MODEL

The model: identical, rented, available to all

The base layer is rented and identical. Everything above it, the rules, the exceptions, the sign-offs, is the only part that produces a different outcome. Firm C has a model and nothing to put it into. Same investment, no return. And three people know how it actually works: when they leave, the firm starts from zero, the next hire takes six months to reach the same judgment, and errors compound in the gap.

Available to every competitor

Reasoning, language, code, speed, scale

Yours alone, if written down

Your rules, your exceptions, your sign-offs, your proof

What you lose when it is not written down

The exception that costs $200K

Written in a 2019 policy memo. Operations and reporting read it differently. Both are defensible. The AI picks one, and the wrong client gets the wrong fee.

The approval that delays a client two weeks

Depends on the amount, the account type, and who is available. Nobody wrote the decision tree, so new hires guess and clients wait.

Which system wins when two disagree

No source of record designated. Each team defaults to its own, and the client gets two different numbers on two different statements.

The definition of correct that lives in four heads

They all agree when asked. None has written it down. When the first one leaves, the definition changes and nobody notices.

A better model does not fix this. These are decisions the firm has never made explicitly. Until someone makes them, writes them down and gets them signed, the firm pays for the ambiguity in errors, delays and rework every quarter.