Four numbers, all from the last year, each printed with who asked, how many they asked, when — and what it doesn’t prove. That last column is the one nobody else prints.
The evidence, with sample sizes and what it does not prove5 rows
| What it found | Who asked | Sample | Fielded | What it does not prove |
|---|---|---|---|---|
| 42% of firms had scrapped most of their AI initiatives, up from 17% the year before | S&P Global Market Intelligence, Voice of the Enterprise | 1,000+ respondents, North America and Europe | October 2025 | Self-reported. "Scrapped most initiatives" is not the same as "AI does not work" — the average organization also scrapped 46% of proofs-of-concept before production, which is roughly what a healthy pipeline looks like. |
| Only 7% of senior leaders can report established ROI on AI | KPMG Global AI Pulse, Q2 2026 | 2,145 respondents across 20 countries | Fielded April–May 2026 | A survey of leaders, not an audit of accounts. Also from the same wave: only 26% have real-time visibility into what their AI costs to run, which is arguably the more useful number. |
| 74% of enterprises rolled back or shut down a live AI customer agent — 81% among those with fully mature guardrails | Sinch, The AI Production Paradox | 2,527 senior decision makers, 10 countries | Fielded January–February 2026 | Vendor-sponsored research by a company that sells the governance layer, so the framing is theirs. The sample and the fielding are real and the direction matches two independent datasets. |
| 53% of data practitioners name poor data quality a top challenge; trust in data is now the single most-prioritized objective at 83% | dbt Labs, State of Analytics Engineering | 363 practitioners and leaders | Published April 14, 2026 | A practitioner survey, self-selected toward people who already use dbt. The internal contrast is the interesting part: 72% prioritize AI-assisted coding against 24% prioritizing pipeline testing and observability. |
The number I’m not using
You’ll see a different one quoted everywhere: MIT’s 95%. It’s a non-peer-reviewed paper from July 2025 built on a hundred and fifty interviews, there’s no 2026 follow-up, and by now every AI vendor’s landing page has quoted it into wallpaper.
The newer numbers are bigger, cleaner and worse for the industry. Those are the ones on this page. A page that dates every figure doesn’t get to make an exception for the figure that flatters it.
What actually killed them
None of those projects died because the model was dumb. They died on data nobody had modeled, a workflow nobody had integrated, and an outcome nobody defined before the build started. That’s my opinion and I’ll label it as one.
But it’s the same place the analytics side keeps arriving from the other direction, which is why the fourth row up there is a survey of data practitioners rather than an AI survey at all. Everybody is standing around the same hole. That hole is the whole of the analytics channel, if you’d rather start there.
If you’re reading this because yours is one of them — the pilot demoed fine and then stopped — the door is the Second Opinion, $2,000 flat, back in 5 business days. Somebody with no stake in the original decision reads it against the three named causes of rollout failure and writes you a page. You keep the memo whoever you end up hiring, including not me.
The three things this studio will actually build, and what each one costs, are on the AI & data page. What it will not build is on the same page, deliberately above the prices.