How Restaurants Use AI for Demand Forecasting (With a Worked Example)

How do restaurants use AI for demand forecasting? They give software their POS history, broken out by day of week and daypart (breakfast, lunch, dinner), along with holidays, weather, local events and reservations, and it predicts how many covers they'll serve and how many of each item they'll sell, shift by shift. Managers turn that forecast into prep lists, purchase orders and schedules, then check it against what actually happened. Done well, it means less food in the bin, fewer dishes running out mid-service and fewer idle labor hours. Big chains already do this every day. For an independent, the real question is whether AI beats the simple methods you can run in a spreadsheet.

The decision this page helps with is whether an AI forecast is worth paying for at your restaurant, which comes down to one test: does it beat a simple method on your own data? Forecasting sits under most of the cost-side uses in our guide to AI for restaurants, because ordering, prep and scheduling all start with a guess about how busy you'll be.

What goes into a restaurant demand forecast

A forecast is only as good as what you feed it. These are the inputs, roughly in order of how much they usually matter.

A study of two food businesses described what this data looks like in practice: multiple strong seasonalities, trend changes, data gaps and outliers. That mess is why forecasting by gut works on an ordinary Tuesday and fails on the days that matter.

What comes out: covers, items and prep by daypart

The output has to match the decision someone makes with it. A daily sales number helps the owner plan cash. It doesn't tell the prep cook how many bowls of chicken to marinate for lunch.

Forecasting methods, from simple to machine learning

AI is the top of a ladder, not a different thing. Each rung adds data and effort, and each has to beat the one below to be worth it.

More complex isn't automatically better. In a 2022 study that tested many models on real sales from a mid-sized restaurant, linear models gave the best one-day forecasts, with a symmetric MAPE of 19.6%. Neural networks did better only on forecasts a week ahead, with a best result of 19.5%. A separate study built forecasts for a casual restaurant and a large staff canteen solely from POS data, in a form a person can read and question. The lesson for an operator: a forecast you can explain and check beats a black box you can't.

What AI adds over a spreadsheet, and what it doesn't

A good spreadsheet forecast is a moving average by weekday, adjusted by a manager for weather and events. AI adds four things:

• More inputs at once. It weighs weather, events, holidays, promotions and recent trend together instead of one adjustment at a time.
• Effects learned from your history. It measures how much rain or a home game moved your sales before, instead of relying on a guess.
• Every item, every daypart, every day. A manager can forecast covers. Nobody hand-forecasts dozens of items across three dayparts each morning.
• Daily retraining. Each day's actual sales correct the next forecast.

It doesn't know what isn't in the data: a new menu item, a road closure, a competitor opening across the street, a private party booked by phone. A manager still overrides for those, and should log each override so you can see later whether it helped.

Large chains already run this way. In Deloitte's survey of restaurant executives, more than a third of them from organizations with 2,000 or more locations, 55% reported daily use of AI in inventory management, with forecasting a common use.

Worked example: from Friday's forecast to the prep list and the schedule

Worked example: The numbers in this section are made up to show the method. Say a fast-casual restaurant wants to plan Friday dinner, 5 to 9 p.m.

Step 1: test two simple methods on the last four Fridays. Its last eight Friday dinners served 176, 184, 170, 190, 182, 196, 168 and 188 covers. Here's how two simple methods would have done on the last four.

A miss is actual minus forecast, so a positive miss means the forecast was too low. The four-week average missed by 5.1% on average and barely leaned either way. That's the bar any AI forecast has to clear on this restaurant's data.

Step 2: forecast this Friday. The four-week average of weeks 5 to 8 is 183.5 covers. Say the restaurant's history shows Fridays with a home game at the nearby high school run about 10% busier, and there's one this week: 183.5 × 1.10 ≈ 202 covers. Learning that 10% from history, instead of guessing it, is exactly what an AI model does.

Step 3: turn covers into items. Say the chicken bowl is 15% of dinner entrées and guests order about one entrée each: 202 × 0.15 ≈ 30 bowls. At 5 oz of chicken per bowl, that's 150 oz, or about 9.4 lb to prep.

Step 4: prep to the cost of being wrong. The bowl sells for $14 and costs $4.20 in food. Running out loses $9.80 of margin per bowl if the guest doesn't order something else instead, plus a disappointed guest. Over-prepping wastes about $4.20 per bowl if the leftover chicken can't be used the next day. Because running out costs more, prep a little above the forecast. The share of nights you want to cover is $9.80 ÷ ($9.80 + $4.20) = 70%.

How many extra bowls that means depends on how the forecast's misses are spread, so the buffer rests on an assumption. Say this restaurant's bowl forecasts have had no lean either way, and the misses form a roughly bell-shaped (normal) curve with a standard deviation of about four bowls. On that curve, the level that covers 70% of nights sits about 0.52 standard deviations above the forecast: 0.52 × 4 ≈ 2 bowls. So prep 32, in two batches, and cook the second to the pace of service. If your misses are lopsided, bigger on game nights, or the forecast runs low, the right buffer is different. A more direct way is to list this item's past misses and pick the buffer that would have covered about 7 nights in 10. A percentile doesn't give a fixed number of extra portions on its own; the spread of your own errors does.

Step 5: staff the peak. Say 30% of dinner covers land between 6 and 7 p.m., which is about 61 covers. If one line cook can handle about 25 covers an hour, the peak needs 61 ÷ 25 = 2.4, so three cooks. The hours either side need fewer.

Step 6: measure and record honestly. Friday serves 196 covers, a 3.1% miss. The bowls sold out at 8:45 p.m., so record demand as "32 or more, sold out", not 32. Recording a sell-out as true demand teaches any forecast, human or AI, to under-predict that item.

How accurate is AI demand forecasting? MAPE explained

Most vendors quote accuracy as MAPE, the mean absolute percentage error. For each period, take the miss (actual minus forecast), ignore the sign, divide by the actual, then average across periods. A MAPE of 5% means the forecast was off by 5% on average. The definition and its weak spots are set out in Hyndman and Athanasopoulos's forecasting textbook, which notes that percentage errors are undefined when actual sales are zero and put a heavier penalty on over-forecasts. Three things follow for restaurants:

1. Zero-sale slots break it. A dessert that sells nothing at lunch on Tuesday produces a divide-by-zero. For item counts, use weighted MAPE instead: add up the absolute misses for the week, ignoring the sign, and divide by total units sold. As a formula: sum of |actual − forecast| ÷ sum of actual × 100. The absolute value matters. If you add the signed misses, over- and under-forecasts cancel: in the worked example, the four-week average's misses of +2, +14.5, −16.5 and +4 net to just +4 covers, 0.5% of the 734 served, though it missed by 16.5 in one week. With absolute misses it is 37 ÷ 734 = 5.0%.
2. It quietly favors forecasts that run low. For example, forecast 100 covers: if 50 come in, MAPE counts a 100% error, but if 150 come in, it counts 33%. The miss is 50 covers both times. A model tuned only to lower MAPE drifts toward under-forecasting, and in a kitchen that means running out.
3. Finer is harder. A day's total covers is far easier to forecast than one item in one daypart. When a vendor quotes a number, ask at what level, for which items and over which weeks.

Always read MAPE next to bias, the average miss with its sign. Positive bias means you keep running short. Negative bias means you keep over-prepping. In the worked example, the four-week average had a MAPE of 5.1% and a bias of +1.0 covers, so it was accurate and balanced.

The most useful single test of any AI forecast is to put its MAPE and bias for the same weeks next to two simple methods: same day last week and a four-week average. If it doesn't clearly beat both on your own data, you're paying for a spreadsheet.

What a better forecast changes: prep, ordering, scheduling and waste

• Prep: item counts by daypart become prep lists and batch sizes, set to the cost of running out versus throwing away.
• Ordering: item forecasts times recipe quantities, minus what's on hand, become purchase orders, and fewer emergency runs to the store.
• Scheduling: covers by hour become shifts, with a minimum crew at peak and fewer idle hours either side. The method is in how to reduce labor cost in a restaurant.
• Waste: ReFED counts overproduction as 11.9% of the surplus food from restaurants and foodservice, with catering overproduction another 9%, and recommends technology that improves demand forecasting. The non-AI steps are in how to reduce food waste in a restaurant.

Food and labor are most of your prime cost, so this is where a forecast shows up on the P&L. Work out yours with the prime cost calculator.

Common forecasting pitfalls

1. Forecasting dollars instead of units. A price increase looks like growth. Forecast covers and items, then multiply by price.
2. Dirty POS history. Voids, comps, renamed items and combos that count as two lines all distort the pattern. Clean them before you trust any model.
3. Treating sell-outs as demand. If an item ran out at 7:30, its sales understate what guests wanted.
4. Leaving one-off days in. A buyout, a storm closure or a broken fryer will be repeated next year unless you flag it.
5. Mixing catering into walk-in demand. Catering and large pre-orders are known in advance. Add them as booked demand rather than letting a model guess them.
6. No history for new items. Start a new dish from a similar item's pattern and adjust weekly.
7. Unlogged overrides. If nobody writes down why they changed the number, you can't learn whether the change helped.
8. Judging by MAPE alone. Check bias, and compare against a simple baseline.
9. The wrong level. The kitchen needs item counts by daypart. A daily sales forecast doesn't help at 10 a.m. prep.

Where Kutlerri.ai fits

Kutlerri runs AI agents for restaurant operators as a managed service, and two of them use demand forecasts. The Prime Cost Optimization Agent, in beta, matches shifts to forecast covers and flags overtime before it lands. The Daily Prep & Waste Control Agent is coming soon and isn't available yet; it is planned to forecast how much of each item to prep by day and daypart and flag over-prep before it turns into waste. Whatever you use, hold it to the baseline test above.

For phones, see voice AI for restaurants. For example vendors by category, see AI tools for restaurants: what they do and how to choose.