Key takeaways
- US restaurant sales swing through the year: in 2025 they ran from $75.8 billion in February to $92.8 billion in May[1].
- Labor takes roughly 33 cents of every sales dollar[2], so a better forecast pays off mainly through better schedules.
- Judge a forecasting tool on weekly error by daypart against your own history, not on a vendor demo.
- Forecasts break around events, weather and promotions; feed those in or override them by hand.
What AI forecasting does that a spreadsheet does not
Most managers forecast by taking the same week last year, adjusting for a hunch, and writing the schedule. That works until something changes: a new office opens nearby, a road closes, Easter moves by three weeks, or a heat wave empties the patio. AI forecasting tools look at the same history but weigh many more signals at once, such as day of week, recent trend, holidays, weather, local events, reservations already on the books and online order volume, and they update the forecast as those signals move.
The output is usually a forecast of sales, covers or transactions by day and by 15- or 30-minute interval, which then drives a suggested labor plan. The value is not the forecast itself. It is the shift you did not overstaff on a slow Tuesday and the server you did schedule for a Friday that would otherwise have been short.
33¢
of each sales dollar goes to labor, roughly the same as food[2]
The seasonality you are forecasting against
Restaurant demand moves in waves. National food service and drinking place sales in 2025, not seasonally adjusted, were lowest in February and highest in May[1]. Employment follows: in 2026 restaurant and bar employment peaked at 12.64 million in June, about 637,000 more than in January[3]. Your own curve will differ by concept and location, but every restaurant has one, and the forecast has to capture it.
Source: US Census Bureau, 2025[1]
Show the data
| Period | Value |
|---|---|
| Jan | 76.8 billion dollars |
| Feb | 75.8 billion dollars |
| Mar | 89 billion dollars |
| Apr | 86.5 billion dollars |
| May | 92.8 billion dollars |
| Jun | 88 billion dollars |
| Jul | 88.2 billion dollars |
| Aug | 91.3 billion dollars |
| Sep | 85.2 billion dollars |
| Oct | 88.5 billion dollars |
| Nov | 84.2 billion dollars |
| Dec | 86.6 billion dollars |
Monthly totals also hide the shape inside each week. A brunch place may do most of its weekly sales on Saturday and Sunday; a lunch spot in an office district may barely open on weekends. Good tools learn these patterns from your POS, not from industry averages.
Inputs that actually move the forecast
| Input | Why it matters | Where it comes from |
|---|---|---|
| Two or more years of sales by interval | Captures weekly and seasonal patterns | POS export, ideally by 15-minute interval |
| Reservations on the books | Shows demand already committed for the next days | Reservation platform |
| Online and third-party orders | Off-premises demand behaves differently from dine-in | Ordering platforms, POS order source |
| Holidays and school calendars | Moving holidays shift demand between weeks | Public calendars, local school district |
| Weather forecast | Rain and heat change patio and delivery demand | Weather data feed in the tool |
| Local events | Games, concerts, conventions, road closures | Manager input, venue calendars |
| Promotions and menu changes | A limited-time offer can lift one item and one daypart | Marketing calendar |
Most tools take the first three automatically once connected. The last two depend on managers telling the system what is coming. If nobody enters the street festival, the forecast will not know about it, however clever the model.
From sales forecast to labor plan
Forecast demand by interval
Start from forecast covers or transactions per 15 or 30 minutes, not daily sales. Labor follows the rush, not the total.
Set labor standards per role
Decide how many covers one server, one line cook or one cashier can handle per interval at your service standard. Write them down by role and station.
Add fixed positions
Some roles do not flex: the opening manager, the dishwasher, the host on a reservation night. Put them in first.
Generate the draft schedule
Let the tool convert demand into suggested headcount per interval, then into shifts that respect availability, minimum shift lengths and labor law.
Manager review
Adjust for what the model cannot know: a new cook who is slower, a private party, a server who is great on the patio.
Compare after the week
Check forecast against actual sales and scheduled against needed hours, by daypart. Feed the corrections back.
The labor standard is the step most often skipped, and it is where most savings come from. Restaurant employees average about 25.1 hours a week[4], so schedules are built from many short shifts; small timing errors at the start and end of each shift add up quickly.
How accurate is good enough?
Ask vendors for forecast error measured as the average percentage miss per day and per daypart, tested on your own history. A tool that is close on weekly totals but wrong on Friday dinner by a wide margin is not helping you schedule. As a working rule, compare the tool's error with your current method over the same eight weeks. If it is not clearly better on your busiest dayparts, the rest of the feature list does not matter.
- Test on a holdout: give the vendor history up to a date and compare its forecast for the following weeks with what actually happened.
- Look at the worst days, not the average. One badly understaffed Saturday costs more than several slightly overstaffed Mondays.
- Check how the tool handles holidays that move, such as Easter and Thanksgiving week.
- Ask how quickly it adapts after a structural change, such as new hours or a remodel.
Forecast dine-in and off-premises separately
Takeout, delivery and dine-in behave differently. Delivery rises on rainy evenings when the dining room empties; dine-in jumps on a sunny Friday when the patio opens; catering orders arrive in lumps that have nothing to do with the weather. A single sales forecast blends these and hides the staffing question that matters: who is needed where. Ask whether the tool forecasts each channel and maps each one to the stations it loads, such as the expo and packing station for delivery and servers for dine-in.
The same applies to prep. A forecast by item and channel can feed prep lists, which matters most for items with short holding times. Start with the ten items that drive most waste and compare forecast-driven prep with what the kitchen did before.
Where AI forecasts go wrong
- Short history: a restaurant open less than a year gives the model little seasonality to learn from.
- Dirty data: comps, voids and test transactions left in the POS distort the baseline.
- Unlogged events: a closed day for a private buyout looks like a demand collapse unless it is flagged.
- Promotion blind spots: a two-for-one week lifts traffic the model will try to repeat next year.
- Self-fulfilling cuts: if the schedule is always lean on Sunday nights and service suffers, sales fall and the model learns to schedule even leaner.
Rule of thumb: every override a manager makes should come with a one-line reason. After a month, those notes show you which inputs the tool is missing.
Why operators are investing, and why many are cautious
Labor scheduling is one of the more mature uses of automation in restaurants. In National Restaurant Association research, 37% of operators planned to invest in automated labor management, recruitment or scheduling systems[5], and 47.6% expect technology, automation, data and AI to improve labor productivity[6]. The pressure is wages: average hourly earnings in food services and drinking places reached 22.25 dollars in September 2026, up from 20.69 dollars two years earlier[7]. A schedule that matches demand more closely is one of the few levers that does not touch the guest experience.
Example: a cafe's Saturday
Example (fictional): a 60-seat brunch cafe always scheduled six front-of-house staff from 8 am to 3 pm on Saturdays. The forecasting tool, fed two years of 15-minute POS data, showed the rush ran from 9:30 am to 1 pm, with the first hour and last 90 minutes at less than half that level. The manager moved two shifts to start at 9:15 am and end at 1:45 pm, and one to start at 10 am. Saturday front-of-house hours fell by about 9 a week with no change in ticket times. In the same example, the tool missed a farmers market opening day; the manager now enters market dates by hand.
Questions to ask before you buy
- Which POS, reservation and ordering systems does it connect to, and how often does data sync?
- Can it forecast by 15-minute interval and by channel (dine-in, takeout, delivery)?
- Will you run a holdout test on our history and show error by daypart?
- How do managers add events and overrides, and are those reasons stored?
- Does scheduling respect our state's predictive scheduling, break and minor labor rules?
- What happens to our data if we leave?
Forecasting sits on top of clean data from your other systems, so read connecting POS, reservations and ordering first, and track the results in a weekly restaurant dashboard. For a sense of where other AI tools stand, see what drive-thru voice AI pilots show and the vendor checks in AI and guest privacy.
Frequently asked questions
How much history does an AI forecasting tool need?
Two years of POS data by interval is ideal because it shows each season twice. With less than a year, expect more manual overrides until the model has seen your full cycle.
Can AI forecasting replace the manager's schedule?
No. It produces a draft that managers adjust for people, events and anything the data does not show. The best results come from managers reviewing the draft and logging their reasons for changes.
What is a good forecast accuracy target?
Compare against your current method on the same weeks. Focus on error during your busiest dayparts, where under- or overstaffing costs the most.
Does forecasting help with food ordering too?
Yes. The same demand forecast by item can drive prep lists and purchasing, which helps with waste. Check whether the tool forecasts at item level or only sales totals.
References
US Census Bureau (2025). Monthly Retail Trade: Estimates of Monthly Retail and Food Services Sales by Kind of Business (1992-present) (opens in a new tab) ↑
National Restaurant Association (2026). Restaurants remain resilient despite challenging business conditions (opens in a new tab) ↑
US Bureau of Labor Statistics (2026). Current Employment Statistics: series CEU7072200001 (opens in a new tab) ↑
US Bureau of Labor Statistics (2026). Current Employment Statistics: series CES7072200002 (opens in a new tab) ↑
National Restaurant Association (2024). Where operators plan to invest in tech (opens in a new tab) ↑
National Restaurant Association (2026). Research Insight: The Future of Restaurant Technology (press release, November 5, 2026) (opens in a new tab) ↑
US Bureau of Labor Statistics (2026). Current Employment Statistics: series CES7072200003 (opens in a new tab) ↑