AI in a spreadsheet: analyse data without guessing
Check duplicate rows, definitions and arithmetic before using AI to analyse a spreadsheet.
Article language: English

In this article
A spreadsheet has six rows, but that does not necessarily mean six deliveries. Before asking AI for the average completion time, establish what a row represents and whether the same number appears twice. Otherwise, you may get a very clear chart of a faulty assumption.
The following data describes an entirely fictional small transport company. CSV is plain text representing a table; here, semicolons separate the columns. Copy it into a spreadsheet or paste it into a conversation:
id_dostawy;status;czas_dni;wartosc_zl
D01;zakonczona;2;100
D02;zakonczona;4;200
D02;zakonczona;4;200
D03;anulowana;;150
D04;zakonczona;3;300
D05;w_toku;;250
In this exercise, find the mean time for unique completed deliveries. The final column is an additional data check. We do not classify it as revenue or proof of payment: the data is insufficient for accounting classification, and the statuses do not confirm payment.
Define the question before calculating
Use a prompt like this:
Work only with the fictional CSV provided. One record should describe one delivery.
Check for duplicate id_dostawy values and list them before calculating.
Do not change the original data. Show the proposed analysis set separately.
Calculate the mean czas_dni only for unique deliveries with status zakonczona.
A missing time is not zero; do not fill it in by guessing.
Show the identifiers used, numerator, denominator and result.
Also sum wartosc_zl for the same deliveries.
Do not infer the quality of the whole company from this small sample.
A correct reference answer detects the repeated D02. The average includes D01, D02 and D04: their times total 9 days, divided by 3 deliveries, for an average of 3 days. Their values total PLN 600. D03 and D05 are excluded, and empty cells do not become zeros.
This is the arithmetic result of a made-up exercise, not a report of a model run. Calculating such a small sample yourself gives you a benchmark for evaluating the answer.
Prepare a file someone can understand
The ChatGPT data analysis documentation recommends clear column names and one record per row. When the tool uses Python, OpenAI recommends checking the code, result and assumptions. You do not need to know how to program to ask for the records used and the calculation.
Add a description of the columns to a real file. Does the amount include tax? Is the date the order date, dispatch date or receipt date? Does a blank mean unknown, not applicable, or a case still in progress? The colour of a header cannot answer those questions.
Keep the original and work on a copy. Remove unnecessary personal identifiers and sheets outside the task. If data comes from documents, first use a controlled PDF reading and terms check (in Polish). A transcription error can survive even a perfectly written formula.
Do not let cleaning change the meaning
In this exercise the exact D02 duplicate is unambiguous. In real data, two entries with the same number might represent two parcels in one delivery. Automatically deleting the second would destroy information. Ask someone who knows the system to define a record.
The same applies to outliers, values far from the rest. A 30-day delivery might be a mistake or an important problem. Do not tell AI to remove it so the report “looks normal”. Record the reason for each change and the number of rows excluded.
Microsoft's Copilot in Excel help warns that AI results can be wrong or misinterpret information. Microsoft advises against using Copilot in Excel for decisions in sensitive areas such as finance, law and medicine. This limitation concerns that product, not every AI tool.
Deliberately change the duplicate problem. In a new fictional version, the second D02 row has a time of 5 days and a value of PLN 220. The first still has 4 days and PLN 200. They are no longer identical. Without knowledge of the system, you cannot tell whether this is a correction, a second part of the delivery or an error.
Assess the two D02 versions before recalculating the average:
D02;zakonczona;4;200
D02;zakonczona;5;220
List the fields that differ. Do not delete either version or select the newer one:
no update dates or change-recording rule have been supplied.
State the question for the data owner and the calculations to hold back.
Prepare a table of cases requiring clarification.
A correct result holds back the average and total including D02 until the discrepancy is resolved. It may identify other complete entries, but should not call their result the result for the whole dataset. The data owner receives a specific pair to resolve, not a request to “clean up Excel”.
Apply safe data-handling practices (in Polish) to that request too. You do not always need the whole customer database to explain a discrepancy. Identify the minimum scope that actually establishes what the disputed rows mean.
Check the arithmetic before the narrative
For a larger file, calculate basic totals or record counts independently in the spreadsheet. Compare how many rows entered the analysis, how many were omitted and why. Select edge cases: a missing date, a return, a duplicate and a record at the end of the period.
Then read the narrative. An average across three deliveries does not justify “the company always delivers in three days”. The guide to checking AI answers (in Polish) develops the distinction between data and inference. The OpenAI Academy data analysis material starts with the decision the analysis should support. Define it before making another chart.
For a final test, use a flawed description. Return to the original CSV, where D02 is an exact duplicate, without the five-day version. A fictional report author wrote: “We completed six deliveries, with an average time of three days. The value of all services was PLN 1,200.”
Check the report's three statements against the original CSV in the article.
Assess separately the number of unique completed deliveries, average time,
and total value of those same completed deliveries.
For each statement show the identifiers and calculation used.
Propose a corrected description. Do not classify the sum as revenue or payment:
the file lacks a basis for accounting classification and information about payment.
A good check retains the three-day average, but corrects the delivery count to 3 and their total value to PLN 600. A true average does not confirm the other statements. Review each claim, especially when some of them look familiar.
If the result will later go automatically to a report or operational system, plan human approval of the action (in Polish). Show the row-selection rule and identified gaps for approval too. A large number on a screen says little about how it was produced or which decision it supports.
Before passing the result on, read the headings and units. “Average time for completed deliveries, in days” defines the scope better than “result”. Retain the information about the exact duplicate removed and excluded statuses. A person receiving the table without the AI conversation should still understand why the calculation contains three deliveries despite six input rows.
When choosing a service for this work, use the AI tool selection criteria (in Polish). In the programme (in Polish) of the AI Course (in Polish), we connect analysis with explaining the result to someone who did not see the conversation; this forms part of the Aeneia approach. In your next report, add one sentence on which records were included. That line often says more than the title of a chart.
*The source states that references were checked on 30 October 2026 and that the OpenAI Academy material was published on 10 April 2026. The data and calculations are a fictional exercise.*











