A food distributor told analysts it finished an AI rollout across every warehouse, then filed a document that never mentions it.
United Natural Foods filed its fourth quarter results on September 8. Strip the markup and the document runs 37,636 characters. Search it. "Artificial intelligence" appears zero times. "AI" as a standalone word, zero times. "Automation," in any spelling, zero times. The word "Lean" appears exactly once, inside a sentence about deploying daily management practices to 44 distribution centers. The company's own press release, published the same day on its investor site, carries that identical sentence and the same absence.
Four hours later, on the analyst call, the chief executive said something different. J. Alexander Douglas Jr. told listeners the company had "completed the rollout of our AI-powered supply chain and procurement planning platform to all DCs in our network," and had "added new AI-enabled features on the UNFI Insights platform" so suppliers could assess store-level performance and improve demand planning. Both sentences appear in two independently produced transcripts of the same call. The claim exists. It simply lives in the room where nobody signs anything.
What gets a number and what does not
The filing is not vague about money. Fourth quarter capital expenditure came in at 117 million dollars against 74 million in the prior year quarter. Full year spending was 217 million against 231 million. Cloud implementation costs ran 17 million dollars in the quarter against 1 million a year earlier, and 35 million across the fiscal year against 7 million, a fivefold increase in twelve months. On the call, management added guidance of roughly 300 million dollars in fiscal 2027 for targeted automation and ERP deployment.
Now look for the matching figure on the other side. There isn't one. The filing says "fourth consecutive quarter of year-over-year improvement in fill rates, on-time deliveries, and throughput." That sentence tells you the direction and how many quarters it has held. It does not tell you how much. No percentage, no basis points, no dollars of recovered margin, no reduction in spoilage or in miles driven.
So the shape of the disclosure is this: inputs precise to the million, outcomes precise to the adjective.
The two percent that came from outside the company
There is one percentage anywhere near an effect claim in the entire call. "I wanted to start with fill rate, up 2%." That line belongs to John Heinbockel of Guggenheim Securities. An analyst said it, not the company. Douglas answered without confirming the figure and without offering one of his own: fill rates are improving in conventional and natural at about the same rate, he said, though conventional fill rates are higher. Direction again. No magnitude.
This distinction is worth holding onto, because it is where most AI evidence quietly falls apart. A number spoken by a third party inside a company's own earnings call reads, in a summary or a slide deck, exactly like a number the company stood behind. It is not. The company declined to stand behind it, in public, on the record, when asked directly. That declination is itself information.
Why the filing is the honest document
It would be easy to read the gap as evasion. I think it is the opposite. The 8-K is the document that carries liability. The call is the document that carries narrative. A chief executive can say "AI-powered" on a call because it describes an architecture, not a result. Putting a quantified AI outcome into a filing means defending it, and defending it means having measured it against a baseline. Most companies deploying these systems right now have not done that. They have the invoices. They do not yet have the counterfactual.
This is the pattern worth watching across every industry where someone claims AI is already working. Bluebeam, owned by Nemetschek, reports that 27 percent of surveyed construction and engineering professionals use AI. BauGPT, a construction software startup, reports 4 percent of German building firms. Both for 2026, both for the predicate "use," a factor of nearly seven apart, and both correct for their own population. Bluebeam surveyed more than a thousand professionals worldwide, of whom 150 were German. BauGPT surveyed German building firms including small trades, whose self-assessed digital maturity came in at a school grade of 3.0. The disagreement is not about AI. It is about who got counted.
Read the filing, not the transcript
Both vendors are selling into the market they are measuring, which does not invalidate their numbers but does tell you which direction the error points. Bluebeam's own text says 46 percent of early adopters saved between 500 and 1,000 working hours. A search engine summary of that same text turned it into "up to 1,000 hours" for everyone. An interval belonging to a subset became a ceiling belonging to all.
The practical move is to stop treating the press release and the earnings call as evidence about AI at all, and instead read the mandatory filing against a fixed word list, because what a company declines to write down under liability is a sharper signal than anything it chooses to say out loud.
That method took four hours of reading down to ten minutes: download the document by its identifier, strip the markup, count against the list, then read only the hits. It produces a negative finding that skimming never produces.
UNFI spent five times more on cloud implementation this year than last and has another 300 million dollars budgeted for automation and ERP in fiscal 2027. That money is real and it is going somewhere. The interesting question is not whether the systems work. It is when the first distributor decides the effect is large enough, and measured well enough, to write the number into the document that counts.