Inteligencia artificial aplicada a menu: the 2026 numbers and what they actually change in your margin

Which food cost figure should feed any pricing engine?

The starting number is 32.0% of sales, the median full-service food cost for 2024 according to the National Restaurant Association, alongside 32.4% in limited service from the same source.

Everything else falls into place under that ceiling. The National Restaurant Association puts the optimal band between 28% and 35%, then splits it by concept: QSR runs 25% to 30%, casual 30% to 34%, fine dining 34% to 40%. An AI engine repricing your menu without knowing which of those four lanes your house plays in will recommend cuts exactly where you need increases. The operating consequence is concrete and dull: before you plug in any software, load the real per-dish food cost using ingredient prices from the current quarter, not from last year. If your figure lives outside that 28-35% band, no algorithm fixes the problem.

Venue size changes the arithmetic before the model does

Two identical houses with different volume cannot hold the same price, and the gap has been measured: the National Restaurant Association reports 33.7% food cost on sales in full service below USD 2 million a year, against 31.0% in houses at USD 2 million or more, during 2024. That is 2.7 points of gross margin the small operator loses purely on buying power and inventory turns. On USD 900,000 of annual revenue, those 2.7 points come to roughly USD 24,300 that never reach the register. A recommendation engine trained on large-chain benchmarks will propose price structures your real cost cannot carry. My criterion, after twenty years sitting in those board meetings: AI helps you buy better and buy earlier, never to paper over a structural scale disadvantage that gets solved in supplier negotiation. The USDA documents food-away-from-home price index increases above 4% per year in 2024 and 2025, and that is where the quiet trap sits.

Repricing late costs more than any software license

On a dish carrying 30% food cost with a frozen menu, two years of that inflation eat between 2 and 3 points of margin without a single alarm firing in the monthly P&L, because sales hold steady while contribution per unit collapses. Consider what happens if your pricing engine runs on 2023 recipe cards: it believes the hake still costs what it cost back then, it models elasticity on a false cost, it recommends holding price to protect volume, you comply, and by year-end profit has evaporated dish by dish while the dashboard glowed green. The bias is systematic and always downward. Refresh recipe cards every ninety days; that calendar is worth more than the AI vendor. Pizza sustains a food cost between 15% and 20% of menu price according to Sauce's 2025 analysis of profitable restaurant foods, roughly half the 32.0% median the National Restaurant Association reports for full service.

Where the real margin lives: dish geometry rules?

No artificial intelligence creates that spread: flour creates it, short labor creates it, and a menu built around cheap inputs with high perceived value creates it.

Here the machine does win, though at a job different from the one being sold to you: sorting your forty dishes by absolute contribution margin and by minutes of station occupancy, crossing both, and telling you which five carry the entire operation. That sorting takes a head chef two weeks by hand; a well-fed model settles it in an afternoon. The blunt takeaway from this block: AI classifies beautifully and decides terribly. Pulling a dish off the menu is still your call. Seventy percent of Americans want more protein, nearly twenty points more than three years ago, according to the International Food Information Council's 2025 Food & Health Survey; and 49% plan to drink less alcohol in 2025, up 44% from 2023 according to NCSolutions.

Demand shifted and your menu probably did not

Read those two figures together and the cash problem shows up: protein is your most expensive input and alcohol was your cleanest margin. Circana also measured a 3% year-over-year rise in consumer spending on food and beverages during the first half of 2025, so the money is there, it simply moved pockets inside your average check. A menu engine fed with the 2022 sales mix has no way to see that turn. The decision these three figures trigger: redesign the non-alcoholic drinks section at cocktail prices before you touch a single plate. Switching suppliers was the number one strategy against rising costs for 40% of operators in 2024, according to TouchBistro. Four out of ten picked the least glamorous, fastest lever, and they were right. Diego F. Parra insists on the correct sequence during Masterestaurant audits, because sequence is what decides the outcome: accurate recipe cards first, purchasing negotiation second, menu engineering third, and only at the end the software that automates what you already understand.

Buying better still beats predicting better

Paying for a recommendation engine while you buy badly means paying to measure waste you could have deleted. I got this wrong for years, recommending technology ahead of purchasing discipline, and results stayed mediocre until I flipped the order. One point of negotiated discount on proteins moves the margin further than any dish description a model will ever write. Cash flow is the leading cause of financial stress and closure among small businesses, according to Inc., and that is where the tension almost nobody resolves lives. A restaurant can hold impeccable food cost inside the 28-35% band the National Restaurant Association marks and still close, because margin is an accounting photograph while cash is a weekly film. AI applied to menus optimizes the photograph. The bridge between them is turnover: a dish at 34% food cost selling eighty units a day throws off more cash than one at 22% selling nine, even though the second looks better on the dashboard.

The cash flow paradox: healthy margin, dead business

When the model tells you to push the dish with the prettier percentage, ask it about units sold and days of immobilized inventory. If it cannot answer, you are not optimizing your business, you are optimizing a spreadsheet column. First: 32.0%, the median full-service food cost for 2024 according to the National Restaurant Association. Monday morning action: calculate your own real figure across the last ninety days and write the difference on a sheet of paper, because that gap is your improvement budget, not an opinion. Second: 2.7 points, the distance between the 33.7% of houses under USD 2 million and the 31.0% of those above that volume, both from the same source and the same year. Action: request three fresh quotes on your five highest-spend inputs, which is exactly what the 40% of operators reported by TouchBistro did in 2024. Third: 4% annual inflation in food away from home according to the USDA in 2024 and 2025.

The 3 figures worth tattooing on your arm

Action: put a quarterly repricing date on the calendar and never delegate it to a software vendor. The decisive difference sits in the input data, not the model: a pricing engine fed with two-year-old recipe costings produces recommendations biased systematically downward, because it believes your ingredient still costs what it cost before the last increase. USDA reports food-away-from-home price index gains above 4% annually across 2024 and 2025; on a dish carrying 30% food cost, two years without repricing eats two to three margin points nobody spots in the monthly P&L. The second difference is scope. Inteligencia artificial aplicada a menu solves classification beautifully —star, plow horse, puzzle or dog on the classic menu engineering matrix— and handles strategic judgement badly. The ceviche losing money may be exactly why the critic came back, and no model holds that fact because it never reached the POS.

Where the promise breaks and where it holds?

Third: demand elasticity is local and seasonal, yet most commercial models estimate it from aggregated category data. A pasta dish in an office district carries a different increase than the same plate in a weekend shopping mall.

As Sam Sifton, editor and food critic at The New York Times, has argued, a menu communicates before the guest reads the first price, and reading that context remains the operator's territory. Fourth, and this is where the cash lives: AI is excellent at finding dishes that hurt profitability and terrible at executing their removal. Pulling a dish touches purchasing, mise en place, the costing of everything else, staff training and regular-guest expectation. That slow work decides whether the exercise becomes margin or one more PDF.