The shelf holds two cases and seven loose cans, while the system shows 25 cans. Before calling the difference a shortage, establish the case size, the count boundary and the latest movements. AI can prepare an exception list; it cannot see the physical storeroom and should not independently change stock balances.
This guide is for restaurant owners, kitchen managers and stockkeepers comparing physical quantities with inventory records. The editorial pilot below covers a few packaged products. It is not a purchasing forecast, an accounting audit or a validated loss-prevention system.
1. Define exactly what you are counting
The BCcampus kitchen-management textbook distinguishes a running stock record from a physical count used to verify it. For our pilot, start with five packaged products in one defined location. Five is an editorial choice of scope, not an inventory standard.
Write down the boundary: dry store S1, for example, rather than the whole restaurant. Products issued to the kitchen have left S1 but may remain within the restaurant’s inventory. Do not write them off again merely because they are no longer on the storeroom shelf.
Each item needs an identifier, an exact description and a base unit: can, bottle or kilogram. “Tomatoes” is insufficient when different pack sizes sit together. Keep raw ingredients, prepared components and finished dishes separate in the first pilot; their transformations need additional rules.
2. Establish a pack-size dictionary before counting
The Oregon Department of Education inventory form records package size separately and uses separate lines for different container sizes of the same food. This belongs to a specific US food programme, not a requirement for a Ukrainian restaurant; the useful principle here is an unambiguous unit.
Our proposed item card:
| Field | What to record |
|---|---|
| Item and location | T01, dry store S1 |
| Base unit | One can of a specified size |
| Pack | Six identical cans in a closed case |
| Evidence | Current label or verified specification, with its date |
| Count rule | Record full cases and loose cans separately |
| Uncertainty | Unknown pack size stops the conversion |
Do not ask the model to guess the contents of a “standard case”. If the supplier changes the pack, the old conversion factor no longer applies. Preserve “two cases plus seven cans” alongside the converted quantity so the number can be traced back.
Do not convert mass into volume using an arbitrary factor. Leave opened packs out of the first pilot; later, they need an approved measurement method accounting for the container, not a model’s estimate from a photograph.
3. Agree on one reconciliation time
The GAO physical-inventory guide, published in 2002, explains why movements must align with inventory records and discusses recounts and investigation before adjustments. These are general control practices, not a new restaurant regulation.
Our editorial proposal is to choose a quiet window and mark a cut-off, such as 09:00. Retain the recorded balance for that exact time. Assign someone to record any issue or receipt during counting; do not interrupt necessary food-safety procedures to accommodate a spreadsheet.
If an item was counted at 09:10 and moved between 09:00 and 09:10, the two quantities are not directly comparable. Movement records and an explicit reconciliation to one time are required. If the movements cannot be reconstructed, do not call the result final: repeat the count in a controlled window.
4. A worked example: a difference is not yet an explanation
All figures below are fictional. Assume one T01 case contains six identical sealed cans. Store S1 has these confirmed records up to 09:00:
| Movement | Quantity, cans |
|---|---|
| Opening balance | 30 |
| Receipt: three cases of six | +18 |
| Documented issue to kitchen | −20 |
| Separate documented write-off | −3 |
| Expected balance | 25 |
Under these assumptions, 30 + 18 − 20 − 3 = 25. The issue and write-off are separate transactions; no can appears in both rows.
The physical record at 09:10 is two cases and seven loose cans: 2 × 6 + 7 = 19. In the first version of the example, there were no movements after 09:00. The variance, defined as physical minus recorded, is 19 − 25 = −6 cans. It does not establish theft, spoilage or an individual employee’s mistake.
During the review, the responsible person finds a record for another six cans issued at 08:50, omitted from the extract. If the document is authentic, belongs to T01/S1 and does not duplicate the already-recorded issue, the corrected expected balance is 25 − 6 = 19. The difference is explained by a missing transaction, not a new write-off.
But if those same six cans were issued at 09:05, after the cut-off, the reasoning changes. Add them back to the 09:10 physical count for comparison with 09:00: 19 + 6 = 25 at 09:00. Do not alter the 09:00 balance using an event that happened later.
5. Give AI the review draft, not the decision
Share only an approved extract without unnecessary personal information or confidential supplier terms. Item codes, locations, units, times and references to permitted records are enough. If preparing a CSV, separately check that identifiers and values survive import.
A reusable brief:
Compare the physical count with the records for the specified location and cut-off time. Use only the supplied pack-size dictionary. Return the original quantity, conversion to the base unit, movement references, variance and questions for a person. Do not fill gaps. Do not explain a cause without a supporting document. Separate movements before and after the cut-off. Do not change balances or create write-offs, purchases or accusations.
NIST describes confidently presented false output as a generative-AI risk. Each proposed explanation therefore needs to return to its document; the model’s wording does not replace evidence.
6. Close the exception with a traceable decision
For an unexplained difference, recheck the location, identifier and pack size, then physically recount, preferably using another trained person. Preserve both results. Do not average conflicting counts to obtain a convenient number.
The decision log should contain the item, original variance, documents checked, confirmed cause or “not established” status, responsible person, date and authorised adjustment if needed. An authorised person makes changes under the restaurant’s inventory procedures, then verifies the resulting balance.
During the pilot, record counting, investigation and review time, the number of unresolved items and the model’s own mistakes. Do not turn a reduction in exceptions into a savings claim without separate evidence.
Takeaway: the useful outcome is an explained—or honestly unresolved—quantity at a specified time. This workflow does not determine food fitness, storage limits or inventory valuation. Use an existing approved spreadsheet and AI service; a new paid system is not required. Nadali has not tested this scenario in a restaurant.
Author: Nadali. Prepared using AI. Sources checked on 28 October 2026.
The cover is a generic AI illustration of kitchen recordkeeping, not the sealed-pack count described here or a test result.
