I stopped asking vendors what their AI does. I ask how they measured it.
The question changes everyone's posture around the table. The slide stays up. The number stays visible. The person presenting has to say the thing they wrote in the deck or the thing they left out of it.
The CFO who scrolled past the 60 percent
A CFO at a Munich logistics group, in September, forwarded me a vendor deck before a procurement call. The headline on slide six was "AI-powered routing increases delivery efficiency by 60%." He wanted me to help him sound sharp on the call.
I asked what he thought the 60 percent measured. He said he assumed it was fleet-wide, over a year, on their own data. I asked what the deck actually said about the sample, the baseline, the time window. He scrolled. Nothing. The deck had the number and the brand of the model. The methodology lived somewhere else, maybe nowhere.
He went into the call with one question in his back pocket. "Show me the methodology behind the 60 percent." The vendor said they would send it after. Nobody has, three weeks on. The silence after that question was the answer.
What the Hamburg audit found
A team at the University of Hamburg's Trustworthy AI Lab audited 100 German AI startups and ran 63 expert interviews. Pranav, Hartmann and Lauscher coded every public marketing claim on a seven-dimension codebook with an inter-rater κ of 0.82. The silence my CFO ran into has a receipt now.
79% of startups making quantitative claims (49 of 62) gave no substantiation. None were fully transparent.
The headline numbers are a stacked gap. 60% market AI prominently without owning a proprietary system. 76% disclose no limitations. The three findings are one pattern a procurement team can run on a vendor's homepage in an afternoon with the published codebook.
There is a sharper piece buried in the results. Wrapper-based AI correlates with overstating the AI's role in the product at ρ = +0.51, p < 0.001. And the vendors most willing to let an LLM write their homepage copy sit in the high-washing group at ρ = +0.25, p = 0.011. The marketing style and the product honesty track each other.
The interview arm is the number an exec can price in. 53% of surveyed professionals had already encountered AI-washing. Their trust in "AI-powered" claims sits at a median of 1 against a median of 3 for the never-encountered group. Mann-Whitney U = 190, p = 0.003.
The procurement team around your table is already applying the discount. The audit says the discount is correct.
Ask for the methodology. The silence after that question is the measurement.