Business AI guide
AI Spreadsheet Analysis: Check the Numbers Before You Share
Check AI spreadsheet analysis with a ten-row CSV, computed answer key and review worksheet. Covers duplicates, refunds, dates, formulas and reconciliation.
In brief
- Define the metric and accepted source rows before calculating.
- Use the synthetic CSV and computed answer key to check your method.
- Keep unresolved dates and provisional results visible.
The short answer
Treat AI spreadsheet analysis as a draft until you can reproduce the important numbers from the source rows. Define the business rule, separate uncertain records, check formulas or code, and reconcile the result before sharing it. A plausible chart is not proof that the analysis answered your question.
- Check duplicate IDs and conflicting records.
- Keep refunds signed correctly.
- Resolve ambiguous date formats instead of guessing.
- Reconcile included and excluded rows to the source total.
- Label provisional results and keep a human reviewer.
A dated practice file, not a product benchmark
This exercise was created and its answer key computed on October 15, 2026. All ten records are synthetic. No customer data was used. We did not run ChatGPT data analysis or Copilot in Excel on this file, and this guide makes no claim about what either product gets right, its accuracy or time saved.
The task is deliberately narrow: calculate provisional net transaction value for October 2026 in USD. This is not accounting revenue recognition, a financial forecast or tax advice. A refund is recorded as a negative transaction on its event date; the example does not try to allocate it to the original sale month.
The complete synthetic CSV
The table contains every source row, including the intentional duplicate. Amounts are USD, with two decimal places.
| event_id | event_type | event_date | amount_usd |
|---|---|---|---|
| A100 | sale | 2026-10-01 | 120.00 |
| A101 | sale | 2026-10-02 | 80.00 |
| A101 | sale | 2026-10-02 | 80.00 |
| R100 | refund | 2026-10-03 | -20.00 |
| A102 | sale | 2026-09-30 | 200.00 |
| A103 | sale | 2026-10-04 | 50.00 |
| R101 | refund | 2026-10-05 | -80.00 |
| A104 | sale | 06/10/2026 | 150.00 |
| A105 | sale | 2026-10-06 | 100.00 |
| A106 | sale | 2026-10-07 | 60.00 |
Define the rule before AI spreadsheet analysis
Use a unique event ID for each transaction. A repeated ID is not automatically a duplicate in every real dataset. For this exercise, the two A101 rows are identical and represent one event, so retain one. If the same ID has different amounts or dates, stop and ask which record is authoritative.
Accept only unambiguous ISO dates in the form YYYY-MM-DD for the first pass. Include events dated October 1 through October 31, 2026, inclusive. The value 06/10/2026 might mean June 10 or October 6. Its meaning cannot be recovered from the string alone, so set A104 aside until the data owner confirms it.
Keep refund amounts negative. Net transaction value is the sum of accepted sales and signed refunds. Do not subtract a negative refund a second time. These rules are part of the exercise, not universal rules for business reporting.
A brief you can give a data-analysis tool
Use only the supplied CSV. First list its columns, row count, repeated event IDs and date formats. For the provisional October 2026 calculation, retain one of the identical A101 rows, exclude September events, and quarantine ambiguous dates. Keep refunds negative. Show the accepted IDs, excluded rows with reasons, sales subtotal, signed refund subtotal and provisional net. Show the code or formulas used. Do not guess a date format, invent transactions or call the result final.
This is an example brief, not a security control. Use only a product, account and dataset approved for the task. File access does not establish permission to upload customer, employee or confidential business data to another service.
The computed answer key
The answer key was computed with a Python script using decimal arithmetic, explicit date validation and the rules above. It is an independent arithmetic reference for this exercise, not evidence that a commercial AI product passed a test.
| Check | Expected result |
|---|---|
| Source rows | 10 |
| Accepted unique October rows | 7 |
| Accepted IDs | A100, A101, R100, A103, R101, A105, A106 |
| Sales subtotal | 410.00 USD |
| Signed refunds | -100.00 USD |
| Provisional net | 310.00 USD |
| Raw sum before checks | 740.00 USD |
| Excluded duplicate A101 | 80.00 USD |
| Excluded September A102 | 200.00 USD |
| Quarantined A104 | 150.00 USD |
Reconcile the result, not only the headline total
The raw signed sum is 740.00 USD. Removing one duplicate A101 (80.00), September event A102 (200.00) and unresolved A104 (150.00) gives 310.00 USD. Independently, the accepted sales total 410.00 and the accepted refunds total -100.00, also giving 310.00. Both checks must agree.
The reported result remains provisional because A104 is unresolved. If the owner confirms October 6, include its 150.00 and recompute to 460.00 USD. If the owner confirms June 10, it stays outside October and the result stays 310.00. Neither interpretation should be silently chosen.
An answer of 740.00 has summed every row without applying the task rules. An answer of 510.00 may have subtracted negative refunds from sales. These are diagnostic examples derived from this file, not observed errors by ChatGPT or Copilot.
How to check AI Excel formulas
Ask what each formula means before copying it across a workbook. Check its ranges, criteria, signs and treatment of text dates. Confirm that headers or totals are not included as transactions and that hidden or filtered rows are treated according to the agreed rule.
After resolving duplicates and ambiguous dates in a separate checked table, one simple reconciliation is to sum the accepted signed amounts. For this small example, adding 120 + 80 - 20 + 50 - 80 + 100 + 60 gives 310. Test a refund, an excluded date and a duplicate separately; a correct grand total can conceal two offsetting mistakes.
Preserve the raw sheet. Keep the accepted and excluded records in separate visible areas with reasons. If a formula uses a helper column, inspect that column too. Do not infer correctness from the absence of an Excel error message.
A reproducible formula check on the accepted rows
Create a new sheet named Accepted and copy only A100, A101, R100, A103, R101, A105 and A106, in that order. Put the event ID in column A, event type in B, normalized date in C and signed USD amount in D, with headers in row 1 and seven records in rows 2 through 8. Keep the original CSV unchanged in a separate Raw sheet. This small checked table makes the calculation easy to inspect.
In an English-language Excel installation using commas as separators, =SUM(D2:D8) should return 310. For a separate sales subtotal, =SUMIF(B2:B8,"sale",D2:D8) should return 410. For signed refunds, =SUMIF(B2:B8,"refund",D2:D8) should return -100. Adding those subtotals should reproduce the net. Some locale settings require semicolons instead of commas; use the syntax your spreadsheet expects.
Test the logic by changing one copied input in a disposable test sheet. Increasing A100 from 120 to 121 should increase both sales and net by exactly 1. Changing R100 from -20 to -21 should reduce net by exactly 1. Restore the original values after each test. If the result moves the wrong way or does not move, inspect the range, sign and criteria rather than trusting the displayed total.
These formulas are deliberately simple. They do not perform deduplication or resolve ambiguous dates. Those decisions happened before the accepted table was created and must stay visible in the exclusion log. Do not paste the same formulas over an unchecked raw file and assume the exercise rules have been applied.
Update the result when the data owner resolves A104
Suppose the data owner confirms in writing that A104 means October 6, 2026. Record that correction in a dated change log, preserve the original string, and add a normalized version of A104 to the Accepted sheet as row 9 with amount 150.00. Extend the calculation ranges to row 9. Accepted rows become eight, sales become 560.00, signed refunds remain -100.00 and net becomes 460.00 USD.
The raw-to-filtered check changes too: 740.00 minus the duplicate 80.00 minus September event 200.00 equals 460.00. A104 is no longer an exclusion because the owner resolved the uncertainty. If the owner instead confirms June 10, 2026, record that resolution and keep the event outside October; the original 310.00 result does not change.
A clarification is evidence, not an opportunity for the AI to rewrite the business rule. Save who confirmed the date, when they confirmed it and what changed. Ask the reviewer to inspect the new accepted row, extended formula ranges, updated exclusion log and revised chart label together.
ChatGPT data analysis and Copilot in Excel: documented features
As of October 15, 2026, OpenAI documents ChatGPT data analysis as supporting uploaded data, tables, charts and Python-based analysis. Its help page says to review generated code, outputs and assumptions. A successful upload does not guarantee complete analysis of a complex or poorly structured file. This describes documented capabilities, not a run on our CSV.
As of October 15, 2026, Microsoft documents Copilot in Excel as producing insights such as charts, PivotTables, summaries and outliers, and generating or explaining formulas. Microsoft explicitly says to review, edit and verify what it creates. Feature access and supported workflows depend on the actual product, account and current requirements. We did not test Copilot on this exercise.
OpenAI: Data analysis with ChatGPT
Microsoft: Get data insights with Copilot in Excel
Microsoft: Frequently asked questions about Copilot in Excel
Inspect the chart and the saved output
A chart should say exactly what its values measure. For this exercise, label the result provisional net transaction value, October 2026, USD, with A104 excluded pending date clarification. Do not label it final revenue or imply a trend from a single incomplete month.
Check axis units, date boundaries, signs and the data range behind the chart. Open the saved workbook or exported report and compare its totals with the answer key. An attractive visual can still hide the wrong source range or a missing refund.
A short review worksheet
| Check | Evidence to keep |
|---|---|
| Scope | Question, period, currency and definition of the metric. |
| Source | Original file, column meanings and row count. |
| Records | Accepted IDs and excluded rows with reasons. |
| Method | Reviewed formulas or code and helper columns. |
| Reconciliation | Independent subtotal and raw-to-filtered checks. |
| Unknowns | A104 remains unresolved; result labelled provisional. |
| Output | Saved file and chart match the checked calculation. |
| Release | Named reviewer and approved audience. |
FAQ: Can I trust a correct-looking total?
Not by itself. Two errors can cancel out. Verify the included rows and the business rule as well as the arithmetic. Keep enough evidence that another person can reproduce the result.
FAQ: Should AI decide what an ambiguous date means?
No. Ask the data owner or use an authoritative data dictionary. Until then, separate the record and label any affected result provisional.
FAQ: Does this prove which AI spreadsheet tool is best?
No. This guide provides a reproducible check, not a product comparison. A fair comparison would run the same approved file, instructions and acceptance tests on named product versions and accounts on a recorded date, then publish the actual outputs and limitations.
Make the next test part of a controlled pilot
If this is your team's first analytical AI workflow, start with synthetic or approved public data, keep the manual fallback and give one reviewer responsibility for the finished result. Expand only after the team can explain and reproduce its checks.
Related service disclosure
The Israeli Institute for AI (IIAI) is a professional institute related to AI NEWS IL's owner, Dr. Yaniv Shenhav. The service below is shown as a related service, not independent product advice.
Sources
Videos worth your time
Is Excel Dead? Meet the ChatGPT Data Analyst.
Spoken English transcript reviewed, not full player playback. Demonstrates inspecting workbook sheets, counting customers, joining tables and checking one aggregate back in Excel. Useful for the source-to-calculation workflow, not proof that Excel is obsolete or that AI outputs are reliable. This is a 2024 interface and subscription description, not current access guidance. Later causal and seasonal suggestions are hypotheses that need evidence. The description promotes an Excel course.
Watch on YouTube →