Enterprise Intelligence Governance

Historic capital. No measurable return.

The technology is not the problem. Under randomised and staggered-rollout designs, artificial intelligence raises issues resolved per hour in customer support by fifteen per cent, cuts time on professional writing by forty per cent, and raises completed tasks among software developers by twenty-six per cent. These are causal estimates from large samples inside operating firms.

The problem is that people cannot tell when it has failed. Organisations do not challenge every decision; they challenge the ones that feel uncertain, and that judgement is made by the person advancing the decision. When confidence rises where accuracy falls, challenge is allocated in inverse proportion to need — and the organisation experiences this as good governance, because nothing anywhere generates friction.

What the paper argues

Artificial intelligence is being deployed at a scale without close precedent. The returns have not followed. This paper assembles a pattern in the experimental literature that has not been put together before, names the mechanism that accounts for it, and derives a governance failure from the two.

Thirty-five randomised and quasi-experimental trials of generative-artificial-intelligence assistance were coded on the hardest outcome each of them reports. Measured gains do not decline monotonically as outcome measures harden. They collapse at one band — objective outcomes carrying a real external consequence — and the collapse concentrates in decision tasks, not in creation or learning.

From this the paper develops Enterprise Intelligence Governance: the structures, decision rights and accountability mechanisms through which an organisation determines which intelligence carries authority, which decisions receive challenge and who owns the resulting judgement. Its central prescription is a shift from confidence-triggered to class-triggered challenge — mandatory independent review determined in advance by consequence, reversibility and outcome verifiability.

Five parts

1. 01
   The deployment and the gap

   Historic capital deployment, large task-level effects, and almost no measurable value in firm accounts — plus the three most-quoted failure statistics that are not research findings at all.
2. 02
   The mechanism

   The confidence–competence inversion, why five established literatures do not quite reach it, and three classes of machine inference that fail in materially different ways.
3. 03
   Governance

   Four architectures fixed by the lifecycle of a single inference, and class-triggered challenge — including the classification criterion existing authority matrices omit.
4. 04
   The finance function and capital allocation

   Why the scarce organisational function becomes adjudication rather than analysis, and four changes to investment appraisal that require no new systems.
5. 05
   Propositions and research agenda

   Four testable propositions, each with a direction and an archival proxy, and a plain statement of what would falsify the argument entirely.

Three things a board should be able to answer

For a chief executive or chief financial officer the operative question is not how much the organisation is spending on artificial intelligence. It is that one. Most large organisations cannot presently answer it — and on the argument advanced here, the ones that can will be the ones that convert.