AI mean and median audit: check six response times

Which duration does each record actually measure?

Every row in this exercise measures elapsed minutes from receiving a fictional request to sending its first reply. It is not the time needed to resolve the underlying issue. All six requests already have an observed first-response duration, use one fictional channel and follow the same start-and-stop definition. We are checking supplied durations, not deriving them from timestamps.

The arithmetic mean adds the durations and divides by the number of observations. The median describes the middle position after arranging the values from smallest to largest. Everyday language often calls either an average, so a report should name the actual measure. A correct number placed under the wrong heading still misleads the reader.

Copy this rule card first: “Practice V1; B1–B6; unit is minutes; include all six supplied completed observations; retain original order; make a separate sorted copy.” Fix that selection before calculating. Do not quietly remove an awkward record because the remaining result sounds more appealing. A changed selection answers a changed question.

For a broader workflow, see analysing business data with AI. The deliverable here is narrower: an inspectable note containing two different calculations and decisions on three prepared claims. It is not a service-level assessment, employee evaluation or comparison of competing AI products.

Keep the source order and create a separate sorted copy

Imagine an equipment rental desk practising a report review. Every number below is invented for instruction. These are not observed customer records, measured model responses or results from Anar’s client work. No target response time or industry benchmark is provided. All six observations are complete; there is no blank value inside this practice set.

First preserve the original sequence: 45, 4, 1, 5, 3, 2. Copy it before sorting. The sorted values are 1, 2, 3, 4, 5, 45; their source identifiers in that order are B3, B6, B5, B2, B4, B1. Retaining those identifiers makes it possible to trace either central value back to the original row.

Count six values in each version. Sorting changes position, not membership. B1 must remain present in both versions. Its duration is much longer than the others, but size alone is not evidence that it was recorded incorrectly. Source correction requires evidence and a documented change, rather than an instruction to make the summary look better.

Check the unit as well as the identifiers. A duration of 45 minutes cannot silently become 45 seconds. Likewise, a reply duration cannot be relabelled resolution time. If you later want a different measurement, define and source it separately rather than improving the wording around an incompatible number.

How do you calculate the mean and the even-count median?

Open a calculator and enter 45 + 4 + 1 + 5 + 3 + 2. The sum is 60 minutes. Divide by six observations: the arithmetic mean is 10 minutes. Write the complete calculation, 60 / 6 = 10, next to the source count. A bare answer is less useful to another reviewer than the numerator and denominator that produced it.

For the median, use the sorted copy, not the original sequence. Six is an even count, so there is no single middle observation. The two middle positions are the third and fourth values, 3 and 4. Their arithmetic mean is (3 + 4) / 2 = 3.5 minutes. Do not choose only the third value because it happens to appear first.

With an odd number of observations, the median would instead be the single middle value after sorting. Microsoft’s MEDIAN documentation explains both cases. Our English card deliberately has six records, so the two-middle-value rule matters. It does not become a five-record exercise when a duration looks inconvenient.

Mean, median and maximum answer different questions. None is a substitute label for another. If you move the figures into a visual, the chart and percentage audit illustrates the wider principle of checking a claim against source values. This exercise requires no generated chart or spreadsheet subscription.

Which three statements should you revise or hold?

Create a claim ledger with four headings: proposed sentence, decision, supporting record and corrected wording or open question. The statements below are intentionally flawed teaching examples. They are not quotations from a tested model. We are practising how to evaluate a report, not claiming that a particular assistant made these mistakes.

Your accepted summary should say: “For these six fictional first-response observations, the arithmetic mean is 10 minutes, the median is 3.5 minutes and the maximum is 45 minutes.” This preserves the count, unit and teaching status. It does not say that every person experienced the median or that a real business achieved a particular standard.

The source does not explain why B1 is longer. Do not add an employee mistake, a system outage or a difficult customer as the cause. Those possibilities require further evidence. Correct arithmetic does not turn an invented explanation into a fact. Keep the unexplained long duration visible and record any investigation as a future question, not a completed finding.

Is the median always the better measure?

No. The House of Commons Library’s statistical literacy briefing explains that unusually high or low values can affect the mean and that the median can sometimes describe the centre differently. That does not make the median universally preferable. The purpose of the analysis matters. Our note retains both measures and leaves a management decision outside the exercise.

A median of 3.5 minutes does not erase the 45-minute observation. A mean of 10 does not imply that most values equal 10. You can see that directly from the small source list. This deliberately constructed set does not establish a population rate, a normal customer experience or the cause of any difference in a production dashboard.

If an unanswered request is added later, do not assign it zero minutes merely to complete the column. Unknown and observed zero are different cases. Define the observation cutoff and inclusion rule before tackling that new question. A new source version must not silently replace this six-observation record while keeping the old result label.

An answered request is also not automatically resolved. The chatbot resolution cohort audit checks confirmation and human assistance, which are separate outcomes from time to first reply. Different measures may belong in the same broader review, but they should not be merged into an unnamed success score.

What can you save and reopen within fifteen minutes?

The core exercise needs a note editor and a calculator, not an AI account. On Windows, open Start, type Notepad and create a new document. Add the heading “Mean and median practice V1: fictional data.” Paste the rule card, the original B1–B6 order, the identifier-preserving sorted copy and the calculation table. Keep your claim decisions underneath the source.

Press Ctrl+S, choose Documents and save as mean-median-review-V1.txt. Close the editor. Open that file from the same folder and check that the original sequence still begins with 45, the sorted sequence begins with 1, and the labels 10, 3.5 and 45 remain attached to mean, median and maximum respectively. Fix a missing source row before treating the handoff as complete.

For an optional AI second opinion, give an assistant you already use only the fictional record and this request: “Preserve source IDs and original order. Show a separate sorted copy. Name the count, sum, arithmetic mean, median and maximum separately. Use the two middle values for an even-sized list. Remove no long observation without source evidence. Review the three supplied claims without inventing a cause or business result.” No particular product account or model output was tested for this article.

Compare the response with the human calculation key, not just its confidence. If it reports a median of 3, point back to the third and fourth sorted positions. If the assistant cannot be accessed, write optional AI comparison not performed. Your manually verified note can still be complete. For a later spreadsheet task, the Excel prompt examples are a separate starting point, not a required account workflow here.

A suggested allocation is three minutes for the source, four for calculations, four for claims and four for saving and reopening. It is a planning allowance, not measured completion speed. In my earlier writing about perfection and effectiveness, I described projects delayed while pursuing every detail. Applied here, that principle means finishing one reproducible review before designing a complete reporting system; it is not evidence that I ran this fictional support experiment. The artificial intelligence guide offers broader next steps.