To check an AI assistant’s answer, start with a sales report you can read yourself. Ask a question with an explicit date range, calculate the answer from the report, and compare both the number and the rows the assistant used. A correct number alone is not enough if the assistant cannot show how it got there.
Use five questions before putting the tool into a manager’s routine: a single-day lookup, a range total, a comparison, a highest-day question, and a question about a date missing from the report. The last one matters because a useful assistant should recognize when it lacks the information to answer.
Choose a dated, non-sensitive sample report
Start with a small report that has clear dates, a defined sales measure, and enough rows to check by hand. A sample is safer for an initial evaluation than a report containing guest names, employee details, or payment information. Ask the vendor whether the assistant accepts a sample file, which formats it supports, and whether you can see the data it actually reads.
The following figures are a hypothetical test, not restaurant results or a claim about any product. The report covers one location from April 1 through April 4, 2026. Each row represents one calendar day. The sales measure is completed-order sales before tax and tips. It excludes refunds, discounts, and canceled orders. Fixed operating costs, variable operating costs, and exception costs are not included; these are sales figures, not profit.
| Date | Completed-order sales | Meaning for the test |
|---|---|---|
| April 1, 2026 | $420 | First day in the sample |
| April 2, 2026 | $510 | Highest daily value |
| April 3, 2026 | $390 | Lowest daily value |
| April 4, 2026 | $480 | Last day in the sample |
The table is the entire source for the test. There is no April 5 row. Give the assistant the report and its definition, but keep your expected answers separate until after it responds. If it has access to other reports, ask how to restrict the test to this sample. An answer drawn from a different source is not a valid result for this exercise, even if the number happens to match.
In a real buying decision, check how the report defines a business day. A late-night restaurant may close a business day after midnight, while another report may use calendar dates. Ask which time zone, location, order statuses, and sales measure the assistant uses. Those choices can change the answer without either screen looking obviously wrong.
Ask five questions with expected answers
Run the questions independently against the sample. Include the year, both endpoints for a range, and the report name or source when the interface allows it. Ask the assistant to identify the rows behind each answer. Record its response before comparing it with the worksheet below.
- Single-day lookup: “Using only the sample report, what were completed-order sales on April 3, 2026?” Expected answer: $390, from the April 3 row.
- Range total: “Using only the sample report, what were completed-order sales from April 1 through April 4, 2026, inclusive?” Expected answer: $1,800. The calculation is $420 + $510 + $390 + $480.
- Two-day comparison: “Using only the sample report, how much higher were completed-order sales on April 2 than on April 1, 2026?” Expected answer: $90 higher, because $510 − $420 = $90.
- Highest day: “Using only the sample report, which day had the highest completed-order sales from April 1 through April 4, 2026, inclusive?” Expected answer: April 2, at $510. All four rows must be considered.
- Missing date: “Using only the sample report, what were completed-order sales on April 5, 2026?” Expected answer: the report does not contain that date. The assistant should not turn a missing row into $0 or supply a number from elsewhere.
These questions cover different risks. A single-day lookup checks whether the assistant can find a row. The total checks inclusion of both endpoints. The comparison checks subtraction and direction. The highest-day question checks the whole range. The missing-date question checks whether the assistant can stop instead of guessing.
Do not help it along after an incorrect answer and then count the corrected response as an initial pass. Record the first answer and, separately, whether a follow-up helped it correct itself. That distinction matters when a busy manager might act on the first response.
Compare the answer with the source rows
For each response, check three things: the stated result, the source rows, and the calculation. A response of $1,800 is insufficient if it cites the wrong dates or offers no way to inspect its basis. Conversely, an assistant that says it cannot answer because it cannot read the file has not passed the arithmetic test; the file or connection needs attention first.
Keep a simple record of the prompt, the first answer, the rows cited, and the outcome. For the range total, verify that April 1 and April 4 are both included. For the comparison, check that the assistant subtracts April 1 from April 2 rather than reversing the order. For the missing date, check that it distinguishes “no data in this report” from “no sales.”
Ask the vendor how a manager can open the underlying report from an answer, and whether the assistant shows the report’s date range and refresh time. If it cannot cite individual rows, ask what other audit trail is available. A summary without an inspectable source is harder to use for a consequential decision.
When you move beyond the sample, compare like with like. A restaurant POS system may show an order total while another view shows a different sales measure. Before judging the AI, agree on whether the question concerns completed orders, all orders, refunds, tips, or another field. Also confirm whether the report has been refreshed since the activity occurred.
If the numbers disagree, do not assume the assistant made an arithmetic error. First check that you both used the same report, location, dates, and definition. Then recalculate from the rows. That sequence separates a data-selection problem from a calculation problem and gives the vendor a specific issue to investigate.
Set a decision rule and handle failures
Decide what counts as a pass before running the test. A practical rule for this sample is that all five first answers must match the expected results, use only the sample, and provide an inspectable basis for the four answers based on existing rows. For the fifth, the assistant must say the date is absent rather than inventing a value.
For example, suppose an assistant gets the four numerical questions right but answers “$0” for April 5. Under that rule, it fails the initial test. The manager should not use it as the sole source for a sales decision, because the missing-date response conceals a gap in the report. Ask the vendor to explain and correct the behavior, then rerun the full set against a fresh sample.
If all five answers pass, the decision is narrower: the assistant may be worth evaluating further for report questions under human supervision. This small test does not prove it will handle larger reports, overlapping locations, unusual adjustments, or every future prompt. Add representative cases from your own workflow before expanding its use.
For a failed answer, save the prompt and response, identify the expected rows and calculation, and report the mismatch to the vendor. Until it is resolved, have a manager use the original report directly. For staffing, purchasing, or other consequential decisions, require a person to review the source and approve the action; an AI-generated number should be an input to that review, not the approval itself.
Keep practice separate from service. Use an isolated sample or a vendor-provided practice environment, and confirm that the exercise cannot trigger a real payment, guest message, or courier dispatch. If you later run a live service pilot, define who monitors it, what decisions remain manual, and how staff will return to the normal report if an answer looks wrong.
Use the result in your software comparison
Accuracy is useful only if the workflow is practical for the people using it. Ask who can access the assistant, how it selects a report, what happens when data is unavailable, and how corrections are handled. Find out whether managers can limit access to sensitive information and whether a report can be inspected without relying on the AI’s summary.
Compare the answer workflow with the reporting views you already need. Fleksa’s partner portal provides dashboard, order, menu, and reservation views, including sales and order summaries for a selected period; available views and configuration can vary. Ask which views and any AI-assisted reporting options are available for your setup rather than assuming a particular assistant or export is included.
For a different management question, such as matching labor hours to sales, use a consistent shift definition and check the source data in each system. This guide to comparing labor hours with sales by shift covers that separate workflow. You may still need other tools for payroll or accounting, so ask how the reports you use alongside them will be obtained.
Next, take the five prompts and your pass rule to a vendor conversation. Compare the relevant workflow and current availability with Fleksa, and ask how a manager would inspect the report behind an answer.
Frequently asked questions
Does a correct total prove the assistant is accurate?
No. Check the dates, source rows, and calculation too. A matching total can hide the use of the wrong report or an incomplete explanation.
What should the assistant do when a date is missing?
It should say the report does not contain the requested date. Missing data is not the same as a recorded day with $0 in sales.
Can a manager use the assistant after it passes this test?
A pass supports further evaluation, not unsupervised decisions. Test representative reports and exceptions, and have a manager verify the source before approving consequential actions.
