PRIMARY-SOURCE FINDING
What the source establishes.
RAND reports that, by some estimates, more than 80 percent of AI projects fail.
This is an estimate cited in RAND's 2024 exploratory analysis. RAND interviewed 65 experienced AI practitioners. Projects that only used pretrained language models through prompt engineering were outside the report's scope.
LIFSYS ENGINEERING INTERPRETATION
Why the evidence matters.
Failure risk often starts before model selection. An unclear operating result, the wrong success measure, weak data, missing infrastructure, or poor workflow fit can make a technically capable model irrelevant to operations.
A prototype can work in isolation and still fail when it solves the wrong problem or cannot enter the real workflow.
DECISION IMPLICATION
The operating decision it changes.
Do not authorize a model build from a use-case slogan. Authorize a bounded integration test only after the result, decision owner, data boundary, and rollback path are visible.
AI PROJECT ADMISSION RECORD / result statement, decision owner, workflow boundary, data readiness, success and failure measures, rollback path.