AI Demand Forecasting for Restaurants: How It Works

The Forecasting Problem in Restaurant Operations

Every morning, a sous chef makes dozens of forecasting decisions: how much risotto to prep, how many proteins to portion, what quantities of garnishes to prepare. These decisions, made largely on intuition and experience, determine whether you run out of an item mid-service or throw away unsold prep at the end of the night.

Even experienced sous chefs have significant error rates in these predictions. A 15–20% forecasting error on prep quantities translates directly into food cost overruns or revenue loss from 86'd items. At scale, this is a major financial problem.

How Traditional Forecasting Works (and Where It Fails)

Traditional demand forecasting in restaurants is based on:

• Historical patterns: "We always sell about 40 salmon on Friday nights"
• Reservation counts: more reservations means more prep
• Manager intuition: experience-based adjustments for events, weather, seasons

These inputs are valuable but incomplete. They miss the interaction effects between variables (a holiday weekend with bad weather on a day with a major local event), they do not account for menu mix changes, and they depend entirely on the knowledge of one experienced person — who might call in sick.

What AI Demand Forecasting Does Differently

AI demand forecasting analyses patterns across multiple data sources simultaneously and identifies relationships that human analysis cannot reliably detect:

Historical Sales at Item Level

AI models train on your POS data at the dish level, not just total cover count. They learn that on rainy Saturdays, pasta dishes outsell proteins by 15%. They know that your special changes the average check but does not significantly affect preparation of base menu items. This granularity is impossible to replicate manually.

Reservation and Event Data

Reservation patterns predict not just volume but mix. A corporate dinner reservation of 30 signals different ordering patterns than 30 individual reservations. A party booking with a set menu eliminates variance entirely for that group.

External Signals

Advanced forecasting incorporates weather data (rain reduces walk-ins, extreme heat increases drink sales), local event calendars (concert at the nearby venue), and seasonal patterns specific to your location — not generic industry averages.

Trend Detection

AI models detect when demand for specific items is trending up or down over time — something that is almost impossible to see in weekly data but becomes clear with months of history. If demand for your vegetarian options has been growing steadily for three months, the forecast accounts for that trajectory.

What AI Forecasting Outputs Look Like

For restaurant operators, AI demand forecasting outputs are practical, not theoretical:

• Item-level prep quantities: prep 47 salmon portions, 23 ribeyes, 31 portions of risotto base
• Purchase list adjustments: buy 12kg of salmon this week vs. the usual 9kg (event weekend)
• Staffing suggestions: Friday needs 2 additional line cooks based on 20% above-average revenue forecast
• Ingredient order timing: order perishables Tuesday for weekend delivery, reducing spoilage risk

The Karu Forecasting Approach

Karu's prep list generation uses your restaurant's own historical sales data combined with current reservation counts and day-of-week patterns to generate prep quantities at the station level. This is not generic AI — it is trained on your specific sales history and calibrated to your menu.

As you use the system, it learns. Corrections you make to the suggested quantities ("we actually sell 20% more risotto on Friday than suggested") become part of the model. Over time, the forecasts become more accurate — and the manual adjustments become smaller.