Three "AI projects fail" stats. Not one is a study.
87%, 80%, 85%. You have seen these numbers in every AI keynote and boardroom deck.
I chased all three back to their primary sources. What I found was not fabrication.
It was degradation, and a citation loop where each number borrows its authority from the next.
"87% of data science projects never make it into production."
Credited to → "VentureBeat" (as if it were research)
Real origin
A single VentureBeat Transform 2019 panel. Deborah Leff of IBM said, off the cuff, that "only about 13%" of projects make it. Subtract from 100 and you get the famous 87%.
What's missing
No sample. No methodology. No paper. One executive's verbal estimate on a stage, now cited thousands of times as if it were a measured finding. The units also drifted over the years, from "data science projects" to "AI projects" to "ML models."
STRIP crack
Whose conditions? It was never a study to begin with.
RAND's own sentence reads: "By some estimates, more than 80 percent of AI projects fail." They are citing other estimates, not reporting their own.
What's missing
RAND's actual work was 65 qualitative interviews about why projects fail. It never produced the 80% number. The headlines credit a rigorous-sounding institution with a figure that institution explicitly disclaimed.
STRIP crack
Whose source? RAND is the megaphone, not the origin.
Gartner forecast that through 2022, 85% of AI projects would "deliver erroneous outcomes due to bias in data, algorithms, or the teams responsible for managing them."
What's missing
"Erroneous outcomes" is not "fail." And it was a time-boxed prediction about a future window, not a measurement of what happened. Both the units and the conditions got trimmed away in the retelling.
STRIP crack
Whose units? Erroneous outcomes became total failure.
DistortedGartner CIO forecast, 2018
The worst thing
The finding
The problem is not one bent number. It is that all three cite each other. Each borrows authority from the next, so the pile looks well-sourced while none of it traces to a single clean measurement. Fabrication is rare. Degradation is the industry standard, and it compounds.
This matters most in financial services, where a slide that says "85% of AI projects fail" gets used to kill or fund real budgets.
The fix is not blanket skepticism. It is one habit for numbers that sound clean:
ask whose units, whose conditions, whose subgroup, and whether a primary source exists at all.
The method
Four verdicts
Every claim gets exactly one, judged only against its primary source.
TrueDistortedUnverifiableFalse
The STRIP test for a single viral stat
One question usually cracks a bad number in under two minutes.
Find the primary source. If there isn't one, stop here.
Whose units? "Erroneous outcomes" is not "failure."
Whose conditions? A forecast is not a measurement.
Whose subgroup? A best-case slice is not the average.
Whose authority? A cited estimate is not the citer's finding.