The forecast is not your problem. The Sunday-night count is.

It is close to 10 on a Sunday night in a Lincoln Park kitchen. The line is broken down, the last table is paying, and a sous chef is on a step stool counting cases of chicken thigh by phone flashlight, writing numbers on a clipboard that will get typed into a spreadsheet on Tuesday. That count sets next week's order, the food cost on the profit and loss, and the argument about why margin slipped. It is also the least reliable number in the building.

Food is the second largest line a restaurant controls. Among full-service operators, food and non-alcohol beverage costs ran a median of 32 percent of sales in 2024, and limited-service operators sat at 32.4 percent, per the National Restaurant Association's 2025 Operations Data Abstract. A point or two of that number is the gap between a good year and a flat one, and most of it is decided by how well stock is counted, used, and reordered. Wholesale food prices have not sat still, which makes the count matter more, not less: when the price of every case keeps moving, a stale par level quietly overbuys. That is exactly the work AI is now being pointed at.

Want the four rooms mapped in your own kitchen? That is one working session

Where the food-cost work actually lives

Before automating anything, map the rooms the number moves through. There are four, and each one hands its error down to the next.

What AI for restaurant inventory management actually does

Strip the marketing and the useful version is narrow. It reads sales out of the point of sale and depletes ingredients per ticket, so usage stops being a Tuesday guess. It forecasts par levels from sales history, day of week, season, and known events, then drafts the order against them. It reconciles what was counted against what was sold and flags the gap, the shrink, the over-portioning, the spoilage, the morning after it appears rather than at month-end. Adoption is real, not talk: in Deloitte's 2025 survey of 375 restaurant executives, 55 percent said they use AI in inventory daily and another 25 percent were testing it, second only to customer-facing uses, and 82 percent expected to raise AI spending. What it does not do is count for you. The physical count, the recipe file that says what each dish draws, and the discipline to keep both current stay human work, and they are the inputs everything else runs on.

What to automate first

The order matters more than the tool. Fix the count, then wire usage, then forecast. Bolt automated reordering onto a bad count and you only order the wrong thing faster.

The trap in a clean forecast

A tidy dashboard invites a dangerous assumption: that the forecast is the truth. It is only as good as the count feeding it, and it cannot see the biggest loss in the room. Restaurants and foodservice generated 12.5 million tons of surplus food in 2024, and nearly 70 percent of it was plate waste, food that guests were served and did not finish, per ReFED. Inventory AI touches none of that. It sharpens purchasing and portioning behind the line; menu design and portion size are where plate waste actually lives. Buy the forecast for what it does, not for a waste problem it will never reach.

The order we would run it

When we map a restaurant's operating week, inventory is usually the fourth room we open, after the reservation book, the phone, and the schedule. The sequence we would run is boring on purpose. Week one, fix the count: one method, one sheet, one unit of measure the whole team uses the same way. Then wire depletion to the point of sale so usage is measured, not remembered. Only then turn on forecasting. The operator walks out of that first week with a count they trust and a variance number that finally means something, and the data underneath it stays theirs to keep. Every purchasing decision after that leans on those two things. None of this needs a new platform on day one. Most kitchens already sit on a point of sale that logs every ticket and a supplier who emails invoices, so the first week is about trusting the count, not buying software.

AI for restaurant inventory management is worth running. It is also the last step, not the first. The kitchen that wins with it is the one that fixed its count before it bought the forecast.