United Natural Foods has been rolling out AI inventory planning since 2023, and the deadline just passed.
Twelve of 49 distribution centers. That was the score in March 2026, three years into a rollout that United Natural Foods intended to finish by August 1st. The company sells roughly thirty billion dollars of groceries a year and moves them through those 49 centers. The software in question is Relex Solutions, a planning platform for inventory and replenishment. Not an agent. Not a copilot. A planning system replacing an older planning system, with a target date that has now come and gone.
That last detail is what makes this story unusual. Almost everything written about AI in food distribution is a funding announcement. Anchr, Burnt, Choco, Afresh, GETASAP, FOBOH, GrubMarket: early-stage companies with early-stage claims, none of which can be checked against anything. Canals raised 35 million dollars in May 2026 to automate sales, customer service, accounting, purchasing and receiving for wholesale distributors. Interesting, probably useful, entirely unfalsifiable for the next several years. UNFI gave a date. The next quarterly report will say whether they hit it.
What a three-year timeline tells you about the real clock
Start with the arithmetic, because it is more instructive than the narrative. If UNFI had 12 sites live in March and needed 49 by August, that is 37 conversions in under five months, after managing 12 in three years. The run rate implied by the first three years is roughly four sites a year. The run rate required by the deadline is roughly eight sites a month. Either something changed dramatically in the second half of the fiscal year, or the deadline slipped.
CFO Giorgio Tarditi described the company as being in an early phase of its transformation, which is a careful thing to say about a project that started in 2023. CEO Sandy Douglas framed the benefit in operational terms: customer service, fill rates, inventory management. Nobody claimed autonomy. Nobody described an agent making decisions. The value proposition is that the right cases sit in the right building on the right day, which is the same value proposition grocery distribution has had since the 1960s, now with better forecasting underneath it.
Around the software, UNFI added automation equipment at four sites and a lean operating framework at 34 of the 49 centers. The lean rollout is further along than the planning platform. That ordering is not an accident. Process discipline is cheaper and faster to install than a system that touches every purchase order in the network.
The vendor pool has been telling the wrong story
Anyone tracking AI in food distribution has mostly been tracking venture capital. The pool of named companies in this niche skews heavily toward seed and Series A, which means the available evidence is pitch decks and press releases. Relex is the other side of that market: an established planning vendor with reference customers, selling into brownfield installations where the hard part is not the model but the master data, the exception handling, and the buyer who has run this region by feel for eighteen years.
> The realistic unit of time for AI in an existing distribution network is not a quarter. It is a fiscal year per tranche of sites, and there are usually several tranches.
That gap between the two vendor populations is where most of the disappointment in this sector gets manufactured. A distributor reads about an AI-native forecasting startup, imagines a six-month pilot, and then discovers that its own item master has 40,000 SKUs with inconsistent pack configurations across regions. The startup did not lie. The distributor was reading a greenfield story and living in a brownfield.
Why the missed deadline is worth more than the success case
Here is the uncomfortable part, and it applies to every reader who has ever signed off on a transformation timeline. If UNFI hit August 1st, the story is a competent rollout by a large company, which is mildly encouraging and analytically thin. If UNFI missed it, we get something far more valuable: a public, dated, quantified account of how much longer AI planning rollouts take than their own sponsors expect, at a company with real resources and a competent operating team. The failure teaches more than the success.
The complication is that we do not know yet, and companies are skilled at reframing. A rollout can be declared complete when the software is technically live at all 49 sites while buyers at 20 of them still override its recommendations daily. Adoption and installation are different numbers, and only one of them usually makes the earnings call. Watch for the qualifier, not the headline.
The practical move is to ask any vendor pitching you an AI planning system for the name of a customer who published a rollout date, and then to check what actually happened on that date, because the gap between the promised timeline and the delivered one is the single most useful number in the entire evaluation.
UNFI reports again soon. Either the 49 centers are running Relex or they are not, and either answer is worth more to a distribution executive than the next ten funding rounds in this category combined. The number to remember is not 49. It is three, as in the years it took to get to twelve.