To check whether a restaurant AI assistant handles missing data, give it an incomplete report and ask a question that cannot be answered in full. A useful response should name the missing period, avoid presenting a partial figure as the full total, and tell you what information is needed. Then repeat the question with a complete report to see whether it can answer when the evidence is sufficient.
Make this a buying test, not a test of how well you can rescue the assistant with follow-up prompts. Use the same wording you would expect a busy manager to use, and decide what counts as a pass before reading the answer.
Build a report with one service period missing
Start with a small report where you know exactly what is present and absent. The following example is hypothetical. The figures are made up for an isolated simulation; they are not a restaurant's results or a claim about any product.
Assume a restaurant has two service periods on Tuesday: lunch and dinner. Its manager wants to know Tuesday's total sales. The incomplete report contains lunch sales of $400 but no dinner row. The complete version contains lunch sales of $400 and dinner sales of $300, for a Tuesday total of $700.
For this test, sales means the same sales measure in both rows. The figures are sales amounts, not profit or deposits. They do not include fixed costs, variable costs or exception costs; none is needed to answer the sales-total question. Do not use the $700 figure to infer profit or operating cost.
Label the incomplete input clearly enough that the assistant can detect the gap. For example, say that Tuesday has lunch and dinner service, then provide only the lunch row. Without the expected service periods, the assistant might not know whether the missing row means no dinner service, zero sales or an incomplete report. Those are different situations.
If your normal workflow uses a restaurant POS or another reporting tool, adapt the format to the report a manager actually receives. Keep this exercise separate from live operations: use fictional data in an isolated simulation, with no real payment, guest message or courier dispatch.
Ask the same question twice
Give the assistant the incomplete report and ask: “What were Tuesday's total sales?” Save the full response. Then start a separate conversation or otherwise clear the previous context, provide the complete report and ask exactly the same question.
Keeping the question constant makes the missing period the important difference. If you add a warning such as “please look for missing dinner data” only in the first run, you are testing whether the assistant follows that warning, not whether it notices the gap in an ordinary manager request.
Use the worksheet below to record the two runs. It shows the decision boundary: the assistant may report the known lunch amount in the incomplete case, but it cannot establish Tuesday's total without the dinner amount or another reliable confirmation of that period.
| Input | Question | Acceptable response | Failing response |
|---|---|---|---|
| Lunch: $400; dinner row absent | What were Tuesday's total sales? | Identifies dinner as missing; says $400 is lunch only; asks for dinner sales before giving a full total. | Calls $400 the Tuesday total, invents dinner sales, or treats the absent row as zero without confirmation. |
| Lunch: $400; dinner: $300 | What were Tuesday's total sales? | Gives $700 and makes clear it combines the two periods. | Refuses to calculate despite having both amounts, or gives a total other than $700. |
The first answer does not have to use those exact words. It should make the limitation unmistakable to someone scanning the answer between tasks. “Sales were $400, but data may be incomplete” is weaker than “Lunch was $400; dinner is missing, so I cannot calculate Tuesday's total.” The former leads with a figure that could be mistaken for the answer.
For a second variation, provide a dinner row explicitly marked “$0” rather than leaving it absent. The assistant should distinguish a reported zero from a missing value. If the manager has reason to question the zero, that calls for a separate verification step; the assistant should not quietly change it into a positive number.
Inspect the answer, not just the warning
A warning alone is not enough if the rest of the response still encourages the manager to act on a made-up total. Read the whole answer and check four things:
- Gap: Does it identify dinner as the missing service period, rather than offering a generic warning about possible errors?
- Scope: Does it label $400 as lunch sales, not Tuesday sales?
- Next input: Does it ask for dinner sales or a report that covers dinner?
- Decision: Does it hold back a full-day conclusion until that input is available?
Also look for assumptions tucked into explanations. A response such as “Dinner was probably closed, so Tuesday sales were $400” fails this test. The report says the dinner row is absent; it does not say the restaurant was closed. An estimate of dinner sales also fails if presented as a reported amount. If the assistant offers an estimate, it must not substitute that estimate for the requested actual total.
Check the complete-report answer with the same care. An assistant that always refuses to answer is not meeting the need either. With both rows present, it should add $400 and $300 and report $700. It should not introduce extra conditions that prevent a straightforward calculation.
Then try a manager's likely follow-up: “Can I use this to decide next Tuesday's staffing?” Even a correct sales total does not show labor hours, guest traffic, service timing or the reason for the sales pattern. The answer should not turn one day's sales into a staffing instruction without appropriate context. For a broader comparison of the inputs, see the guide to comparing labor hours with sales by shift.
Set acceptance rules and a human review step
Before a demonstration, write down what the assistant must do and who will review it. For this test, a pass requires all of the following: it flags the absent dinner period, keeps the lunch figure separate from the day total, requests the missing input, and calculates $700 when both rows are supplied. A plausible-sounding response that misses any of those requirements is a fail for this scenario.
Keep the input, question and complete response for each run. Note whether a manager had to prompt again to get the gap acknowledged. A follow-up correction can be useful, but record it separately from the first answer. Otherwise, the record may make the assistant look safer in ordinary use than the initial response suggests.
Repeat the simulation with report formats your team commonly sees. Try an absent row, a blank value and an explicit zero. Also try a report whose period is unclear: the appropriate response may be to ask which dates or services it covers before comparing totals. These variations help you find whether the assistant handles the actual handoff from a report to a manager's question.
Ask vendors how the assistant receives report data, how it indicates incomplete inputs, and whether a manager can inspect the figures behind an answer. Ask what happens when a report is updated after a question is asked. If an export or connection matters to your workflow, confirm that it is available for your setup rather than assuming every reporting tool supplies it.
Keep a person responsible for checking the underlying report before acting on an AI summary. A simulation is a controlled check, not a live service pilot. If the simulation passes and you choose to run a live pilot, define who reviews each answer, which decisions remain with the manager, and how staff report a questionable result. Stop and investigate if the assistant labels partial sales as a complete total.
To take the next step, bring your acceptance worksheet and a sample report format to a product conversation. Review your management software needs and discuss requirements with Fleksa to compare the workflow and current product fit.
Frequently asked questions
Should an AI assistant refuse every incomplete report?
No. It can describe figures that are present if it labels their scope clearly. It should not present a partial amount as a complete answer when the missing information is needed.
Is a missing sales row the same as zero sales?
No. A missing row does not establish what happened during that period. An explicit zero is a reported value; an absent row calls for clarification or the missing report data.
Does one passing test prove the assistant is safe for every decision?
No. Test other report formats and questions your managers use, then keep human review in place. A pass on a sales-total question does not establish that the assistant can make staffing or other operating decisions.
