UNFI set a checkable date for finishing an AI rollout across 49 distribution centres, which makes it rare.
Twelve of forty-nine. That was the number United Natural Foods gave in March 2026, three years into installing a single planning platform across its distribution network. The target was all forty-nine centres by 1 August 2026, the end of its fiscal year. That date has now passed. Whether the company hit it will show up in the next quarterly report, and either answer is more useful than most of what gets written about AI in food distribution.
Almost everything published about this niche describes companies that do not yet have customers. Anchr, Burnt, Choco, Afresh, GETASAP, FOBOH, GrubMarket: seed and Series A, promising, unproven at scale. The funding announcement is the event. Nobody says what happens in month thirty-one of a deployment, because nobody has been there. UNFI has. It runs the largest natural and organic food distribution network in North America, it started this work in 2023, and it is publicly traded, which means it has to keep talking about it whether the news is good or not.
What a three-year timeline actually tells you
The platform is Relex Solutions, a Finnish company that has been selling retail and supply chain planning software since 2005. Not an agent. Not autonomous anything. It forecasts demand and plans replenishment, and the value it creates lands in fill rates and inventory levels, which is to say in working capital and in whether a grocery store gets what it ordered. CEO Sandy Douglas framed the benefit on the 10 March call in exactly those terms: customer service, fill rates, inventory management. CFO Giorgio Tarditi went further and described the company as in an early phase of its transformation. Three years in, with a quarter of the network live, and the CFO calls it early.
Read that against the pitch decks. The gap between "AI reduces food waste by 30 percent" and "we are early, three years in, twelve sites of forty-nine" is not a gap in ambition. It is the difference between a greenfield claim and a brownfield fact. Every distribution centre has its own product mix, its own local demand patterns, its own warehouse management system with its own accumulated exceptions, and its own people who have been overriding the forecast by hand for a decade because the forecast was wrong. You do not swap a planning system into that in a quarter. UNFI is also running automation equipment at four sites and a lean operating framework in 34 of the 49 centres, which suggests the planning platform is one strand of a much larger operational rebuild, and that the strands constrain each other.
The replacement, not the agent
The second thing worth noticing is what UNFI bought. In a year when every distribution vendor is describing autonomous agents that handle sales, customer service, accounting and receiving, and when Canals raised 35 million dollars in May 2026 on exactly that pitch with Base10 Partners leading, the largest visible AI project at a major distributor is a planning system replacement.
That is not a conservative choice. It is a correct reading of where the money is in distribution. Fill rate is the number that determines whether a retail customer renews. Inventory is the number that determines whether the business generates cash. An agent that drafts customer service replies touches neither. A demand forecast that is two points better touches both, on every SKU, in every centre, every day. The unglamorous system sits closer to the P&L than the impressive one.
There is a version of this argument that goes too far, and it is worth naming. Relex is not new, planning software is not new, and it is entirely possible that what UNFI is doing would have been called an ERP project in 2016 and gets called AI in 2026 because that is what the analysts want to hear. The machine learning in demand forecasting is real, but the phase that takes three years is data integration, process redesign and getting planners to stop overriding the output. That phase has nothing to do with the model. It never did.
Why the missed or hit deadline is the story
Which brings us back to the date. Companies almost never give falsifiable timelines on AI deployments, because a falsifiable timeline can be missed, and a missed timeline is a headline. UNFI gave one anyway, in a public call, with a hard number attached. Thirty-seven centres in under five months, after three years to do twelve, would have required the pace to increase by roughly a factor of nine. If they did it, the earlier slowness was a learning curve and the curve flattened, which is the best available evidence that these projects get easier. If they did not, the industry has its first credible number on how long AI rollouts in food distribution actually take, and it is longer than anyone is budgeting for.
The practical move for anyone evaluating a distribution vendor is to stop asking what the system can do and start asking the reference customer how many months passed between contract signature and the first site running without manual overrides. That single question separates the vendors who have crossed the brownfield from the ones who have only modelled it.
Both answers to the UNFI question are worth having. One tells you the ramp is real; the other tells you the planning assumptions across an entire sector are wrong. The next quarterly report settles it. Very few things in this niche settle at all.