Eight portions remain at the end of service. That does not tell you how many to prepare tomorrow. You need to know what sold, whether the dish became unavailable, what changed that day and what happened to the unsold food. AI can help prepare that review; the person running the kitchen should approve the next production plan.
The workflow below is an editorial proposal for a small restaurant or café pilot. It is not a validated forecasting method or a promise of savings. Start with a spreadsheet, a waste log and one dish whose records the team can check.
Separate the reasons food is wasted
UNEP’s Food Waste Index Report 2024 attributes 28% of the food waste generated at retail, food-service and household levels in 2022 to food service. Its overall estimate includes inedible parts. That establishes context, not a savings target for your restaurant.
The WRAP / Guardians of Grub seven-day tracker separates spoilage, preparation waste, plate waste and other waste. Keep those distinctions: peelings, an unfinished customer meal and unsold production call for different questions.
For this pilot, define an additional operational category: prepared portions that remained unsold and were actually discarded. Do not assume every waste entry means overproduction. If the reason is missing, mark it unknown rather than asking AI to invent an explanation.
Build a small, consistent dataset
Choose one dish and one service, such as Tuesday lunch. Gather several comparable weeks. Four matching weekdays can support an initial conversation, but cannot establish seasonality. This is our suggested pilot design, not a measurement standard.
Use one row per dish and service, recording:
- Date, service period and a consistent unit of measurement
- Opening available portions and additional portions prepared during service
- Portions sold, other recorded uses and closing balance
- Unsold discarded portions and reasons, without double counting
- Any time the dish became unavailable, or confirmation it remained available
- Known circumstances such as bookings, promotions, menu changes or shorter hours
Check the balance: opening portions + prepared = sold + other uses + unsold discarded portions + closing portions. The categories on the right must be mutually exclusive; never count a portion as both discarded and closing stock. Record plate waste separately; a sold portion is already counted in sales.
A closing balance is not automatically waste, and an inventory entry does not authorise later use. Staff responsible for food safety determine the food’s status through the restaurant’s established procedures.
Reconcile sales with the till system, handling cancelled items and duplicate records consistently. Record staff meals and complimentary servings separately. Leave missing values visibly missing. If portion sizes changed, resolve that difference before comparing counts.
Only provide an external AI service with an extract your business has approved for sharing. Guest names, phone numbers, payment details and personal notes are unnecessary. Commercial records also need appropriate access and retention arrangements; our guide to data boundaries explains the first checks.
Work through an example before changing production
These are entirely invented figures for four Tuesday lunch services. The example assumes no opening stock, identical portions, no other uses, no stockouts and disposal of all unsold portions.
| Service | Prepared portions | Sold portions | Unsold portions discarded |
|---|---|---|---|
| 1 | 36 | 26 | 10 |
| 2 | 36 | 30 | 6 |
| 3 | 36 | 28 | 8 |
| 4 | 36 | 28 | 8 |
| Total | 144 | 112 | 32 |
Sales range from 26 to 30 portions, with a median of 28. The discarded-unsold share is 32 ÷ 144 × 100 = 22.2%. That measures one dish, not the restaurant’s total food waste.
At an illustrative ingredient cost of UAH 24 per portion, the discarded production represents UAH 768 in ingredients: 32 × 24. This is neither lost sales revenue nor guaranteed recoverable savings. Labour, energy and other costs are excluded.
One proposal for discussion is to test 32 portions instead of 36 at the next comparable service: four fewer portions, an 11.1% reduction in planned production. The number 32 is an editorial test choice, not an AI-discovered optimum. The kitchen manager must consider current bookings, whether further preparation is feasible within approved procedures and the risk of turning customers away.
If the dish previously sold out, recorded sales do not reveal all demand. Do not reduce production solely on that history; first examine periods of unavailability and recorded unfulfilled orders.
Ask AI for a review you can audit
Keep arithmetic in spreadsheet formulas. Ask AI to identify inconsistencies, summarise recurring observations and prepare questions. NIST’s generative AI risk profile identifies confidently expressed false content as a risk. Check the underlying rows rather than trusting a fluent explanation.
A reusable brief:
Review the attached table for one dish. First list missing values, duplicates, inconsistent units and rows that do not reconcile. Do not fill gaps with assumptions. Separate observed facts from possible explanations. Give source rows and formulas for every numerical claim. Compare matching service periods only and flag stockouts. Provide up to three questions for the kitchen manager and one possible production adjustment for their consideration. Do not call it a forecast or guaranteed saving. Do not recommend changes to storage limits, food reuse, recipes or allergens. Do not place orders or change operational systems.
Adapt the wording to your records while keeping the goal, sources and expected output explicit.
Approve one change and measure the trade-offs
Before service, a named person checks the figures, approves the quantity and records the reason. Keep the original data, AI suggestion and approved plan. The assistant should not independently alter purchasing, menus, recipes or production instructions.
Compare similar services using three measures together: discarded-unsold portions per 100 prepared, periods when the dish is unavailable, and time spent preparing and checking the review. Note changing bookings and other demand factors. Lower waste alongside more disappointed customers is not sufficient evidence of success.
Keep decisions manual when the records do not reconcile, the menu has changed substantially or nobody can approve the result. Food suitability, temperatures, storage periods and allergen decisions belong in the restaurant’s existing food-safety process, outside this experiment.
Your next step: pick one dish and see whether you can account for its movement through one service. If not, improve the log first. A useful pilot produces a checked decision, which may be to leave production unchanged.
