Five of the most-quoted numbers in business this year have no traceable origin. Not one of them is a lie. Every one of them is being used to approve capital.
Somebody in your organization is currently building a slide with a percentage on it. The percentage came from an article. The article came from a summary. The summary came from a post that was itself summarizing a press release. Somewhere at the bottom of that chain is a real piece of research, and somewhere in the middle of it, the date fell off.
We spent several weeks this summer checking numbers before publishing them. The failure rate was higher than we expected, and the pattern in the failures is more interesting than any individual error.
Five that do not survive checking
"95 percent of AI pilots fail." This is the most-repeated business statistic of the year. It comes from MIT NANDA's The GenAI Divide, published in August 2025. There is no 2026 successor study. It is now routinely printed under 2026 headlines with no date attached, and in circulation it mutates — we found the same finding rendered as 90 percent, 88 percent and 80 percent, each with no new source.
"Gartner says 89 percent of AI agent pilots never scale." We could not find any Gartner publication containing this claim. Gartner has published several relevant predictions, none of which say this. The number appears to have been generated somewhere in the retelling of a different one.
"86 percent of CIOs are moving workloads back from the cloud." Repatriation figures in the 80s circulate constantly. We could not trace any of them to a primary survey. They appear to be restatements of much older survey lines, gradually rounded upward.
"$690 billion of hyperscaler capital spending in 2026." Widely quoted, sometimes as $725 billion. It appears in no filing from any of the four companies it purports to aggregate. Individual guidance is available and citable — Microsoft's own figure is roughly $190 billion — but the aggregate exists only in secondary blogs.
"$176 billion wiped from tech earnings by depreciation changes." This is an estimate by an individual investor of what would happen if companies shortened asset lives from four-to-six years to two-to-three. That change has not occurred. There is no company disclosure, no auditor finding and no regulatory action behind it. It is a conditional projection being reprinted as an event.
Nobody in that chain lied. Each step was a reasonable summary of the step before it. The chain simply has no integrity check, and provenance is the first thing that falls off when text gets compressed.
Why it got worse, specifically now
Business statistics have always degraded in transmission. What changed is the number of hops and the speed.
A summary of a summary used to require a human to write it, which took hours and imposed a natural limit. It now takes seconds and happens continuously, at scale, in both directions — content generated from summaries, and summaries generated from that content. Each pass compresses. The finding survives compression because it is the interesting part. The methodology, the sample size, the fieldwork date and the distinction between measurement and forecast do not survive, because they are the boring part.
Then a new article is published in 2026 containing a 2025 finding, and the publication date silently becomes the finding's date in every subsequent retelling.
There is a selection effect on top of this, and it is the part worth thinking about strategically. Numbers that survive laundering are the ones that are memorable and useful to whoever repeats them. A precise, hedged, methodologically careful finding does not travel. A round, dramatic, quotable one does. So the pool of "things everyone knows" in any industry drifts steadily toward the quotable and away from the accurate — not through anyone's dishonesty, but through ordinary selection.
Visual 1 — What people think they are citing, and what they are actually citing
The claim | What it is treated as | What it actually is |
|---|---|---|
95% of AI pilots fail | A 2026 finding about current conditions | An August 2025 study, undated in circulation, with no 2026 successor |
89% of agent pilots never scale | A Gartner finding | No such publication located |
86% of CIOs repatriating workloads | A recent survey result | No traceable primary survey |
$690bn hyperscaler capex | An aggregate of company disclosures | A blog aggregation; absent from all four companies' filings |
$176bn depreciation impact | A quantified earnings effect | One investor's conditional estimate of an event that has not happened |
How to read it: Column two is what appears in the meeting. Column three is what is actually behind it. In four of five cases, the gap is not that the number is wrong — it is that it cannot be evaluated at all, which is a different and worse problem.
The trap is set for careful people
Here is the uncomfortable part, and the reason this is a leadership problem rather than a research one.
Notice the direction of those five. Most are pessimistic — pilots fail, projects don't scale, companies are retreating from the cloud, earnings are about to take a hit. They are the numbers a skeptical executive reaches for when pushing back on enthusiasm.
Which means the person most likely to deploy an unverifiable statistic in a meeting is often the person who thinks they are being rigorous. They are resisting hype, they have a figure, and the figure has been laundered clean of everything that would let anyone check it. Sophistication is not protection here. If anything it is the opposite, because a skeptical number faces less scrutiny in a room full of enthusiasts than an optimistic one does.
Three questions, thirty seconds each
Who collected this, and when was the fieldwork? Not when the article was published — when someone actually asked the question. A 2025 survey published in a 2026 article is a 2025 survey.
What was the sample, and who paid for it? Vendor-funded research with disclosed methodology is usable with a caveat. Research with no stated sample is not research.
Is this a measurement or a forecast? "40 percent of projects will be canceled by 2027" is a prediction. It is not evidence about today, and the two get conflated constantly — often in the same sentence.
Any number that cannot survive all three does not belong in a document that authorizes spending.
What to actually change
Require a source and a fieldwork date beside every external number in board material. Not a footnote at the back — beside the number. The requirement does most of the work on its own, because the numbers that cannot satisfy it tend to quietly disappear during drafting.
Label forecasts as forecasts, in the text. Predictions are legitimate inputs. They are not observations, and a deck that mixes them without marking which is which cannot be reasoned about.
Check one number per deck, and say that you do. Not all of them — that is not a realistic use of anyone's time. One, chosen at random, checked to origin. The deterrent effect vastly exceeds the checking effort, because the author never knows which one it will be.
Treat "everyone knows" as a warning. The statistics that reach the status of common knowledge in your industry are, by the selection effect above, disproportionately the ones optimized for repetition rather than accuracy.
Keep your own primary sources. Your customer data, your win-loss records, your churn — these are the only numbers in your business whose provenance you fully control. Most organizations under-use them precisely because they feel less authoritative than a figure with a research firm's name attached.
A decision made on a real number can be wrong, and you will find out why. A decision made on a number with no origin cannot be evaluated at all — not before, and not afterward, when you are trying to work out what you got wrong. That is the actual cost here, and it does not appear until the post-mortem, at which point the slide has been deleted and nobody remembers where the figure came from.
Sources and method. A LookatBusiness original. The five claims examined here were checked in the course of research for this publication during July and August 2026. "95% of AI pilots fail" traces to MIT NANDA, The GenAI Divide: State of AI in Business, August 2025; we located no 2026 successor study. We were unable to locate any Gartner publication containing the claim that 89 percent of AI agent pilots never scale; Gartner has separately published that over 40 percent of agentic AI projects will be canceled by end-2027 (June 25, 2025) and that 40 percent of enterprises will demote or decommission autonomous agents by 2027 (May 26, 2026) — two distinct predictions that are frequently conflated. Cloud repatriation figures in the 80 percent range could not be traced to any primary survey. Aggregate 2026 hyperscaler capital spending figures of $690bn and $725bn appear in no company filing we could identify; Microsoft's approximately $190 billion guidance is company-sourced. The $176 billion figure is an estimate attributed to investor Michael Burry, conditional on a change in depreciation schedules that has not occurred; we found no SEC or auditor action on the point. Where we could not verify a claim, we have said so rather than omitting it.
