UNFI finished rolling out AI planning across its network, and even its own investor communication cannot say what the software did.
Forty-nine distribution centers. One planning rollout, finished. In early September, United Natural Foods reported its fourth quarter and full fiscal year 2026, and the story it told investors was about operational discipline. The AI planning system that now runs across the network was in the room. It just wasn't the thing being credited. The press release names the method, a Lean operating approach introduced alongside the software, rather than the tool itself.
On the surface that's a matter of corporate wording. Underneath it's a measurement problem, and almost every company that buys planning software is about to repeat it. UNFI introduced two interventions into the same warehouses over the same period. One was a new way of working: standard work, visible flow, daily problem solving. The other was a forecasting and planning layer that decides what to order, where to hold it and when to move it. Both aim at the same outcomes, fewer stockouts, less spoilage, steadier labor. When those outcomes improve, there's no clean way back to the cause.
When Two Changes Share One Result
Food distribution makes this especially hard. Margins are thin, volumes swing with seasons and promotions, and a single bad week of produce can erase a quarter of careful work. Any improvement shows up in a handful of metrics that everything touches at once: fill rate, shrink, cases per labor hour. A planning tool improves those numbers. So does a Lean program. So does a mild summer, a renegotiated supplier contract or a new shift manager in one building.
A company with 49 sites could have measured this. It had a natural experiment sitting in its own network. Some centers could have received the planning system first while others kept running the old process with the same Lean training. Comparing the two groups over a few months would have isolated what the software added. Once everything goes live together, that option is gone for good. You can't build a control group afterwards.
This isn't a criticism of UNFI's results. The combined program may well be working. The point is narrower and more useful: a large, well-resourced distributor finished a full AI rollout and, in its main investor communication, chose to credit the method. That's either a deliberate signal about what management believes did the work, or an admission that the separation can't be made. Both readings should interest anyone signing a planning software contract.
Why the Mittelstand Should Care More Than UNFI
A listed company can live with blended attribution. Its investors judge the whole program, and the next budget cycle rarely asks which line item earned its keep. A mid-sized distributor or manufacturer in Germany is in a different position. The planning license is a visible, recurring cost. Within eighteen months somebody, often the owner, will ask whether it paid off. If the honest answer is "things improved, but we also changed how we work," the license renewal becomes a matter of faith.
The pattern repeats because vendors and consultants both prefer bundled rollouts. Software implementations go more smoothly when processes are cleaned up at the same time, and process programs look more modern with a tool attached. Each side gets to claim the improvement. The buyer is left with a result and no way to explain it.
There's also a quieter cost. If the method did most of the work, the company may be paying for software that adds little beyond what disciplined planners with a spreadsheet would achieve. If the software did most of the work, the Lean program may be running on goodwill. Neither question can be answered later. Both could have been answered in advance, cheaply.
The practical move is to decide, before the first site goes live, which locations get the tool and which don't, write down the three numbers you'll compare, and hold that split for long enough to see a full ordering cycle.
That doesn't require a statistics department. Two comparable warehouses, or two product groups in the same warehouse, are enough to tell signal from noise in most cases. The discipline isn't in the analysis. It's in accepting a few weeks of slower rollout in exchange for an answer you can defend.
UNFI will likely keep improving, and nobody will ever know exactly why. For a company of its size that may be an acceptable trade. For a business where one software contract is a meaningful share of the IT budget, it isn't. The question to put to any vendor is simple: how will we know in two weeks whether it was your product? If neither of you has an answer, the rollout plan isn't finished yet.