You do not need to be an analyst to get analyst-grade answers from a spreadsheet anymore — you need to ask clear questions and check the arithmetic. This module covers cleaning messy data, plain-English formulas, Copilot in Excel and Gemini in Sheets, chatting with a CSV, and the reconciliation habit that catches AI's miscounts.
Ritika is a finance analyst at a Gurugram logistics firm, and the monthly sales export just landed: 1,800 rows where dates appear in three formats, region names are spelled four ways ("North", "NORTH", "N.", "north zone"), currency symbols are mixed into the amount column, and forty rows are duplicates from a system glitch. Her manager wants "the story" by Thursday — which regions grew, which products slipped, anything odd. The old Ritika would lose Tuesday to cleanup and Wednesday to pivot tables. The new workflow: twenty minutes of AI-assisted cleaning, an hour of asking questions in plain English, and — this is the part that keeps her employed — a careful pass where she verifies every number the AI computed before a single one enters her memo.
That last pass exists because of the strangest fact in this whole track: the same AI that writes flawless prose is unreliable at arithmetic. Everything in this module is built around that asymmetry.
Almost no real export arrives clean, and cleaning is where AI assistance pays first because the risk is low — you can see whether it worked.
Clean a copy, never the original. Duplicate the sheet or file first, so the raw export survives as your source of truth — you will need it for reconciliation later, and one bad find-and-replace can silently corrupt data you cannot recover.
The formula barrier is gone. Describe what you want in plain English — "sum the Amount column, but only rows where Region is North and the date falls in March" — and any chat tool returns the SUMIFS with your actual column letters if you shared the layout. Two disciplines make this safe:
This works in reverse, too — inherited spreadsheets full of formulas someone else wrote stop being frightening. Paste the formula and ask what it does. Wrapped in that habit, AI turns every cryptic legacy workbook into a documented one.
Both suite AIs can look at your actual data in place, which chat tools cannot do without an upload.
Open the Copilot pane and ask questions about the sheet: "What are the trends in this data?", "Add a column categorizing deals by size", "Show total sales by region as a chart." Copilot works best when your data is formatted as a proper Excel table with headers; it will propose formulas, highlight patterns, and insert pivot-style summaries you approve before they land.
The Gemini side panel in Google Sheets answers questions about the open sheet, generates formulas, and creates summaries and charts on request: "Which product had the biggest month-over-month drop?" It can also generate tables from a description when you are starting fresh.
Their shared limits matter as much as their features. Both struggle when data is messy (merged cells, multiple mini-tables on one sheet, headers mid-sheet) — clean first, then ask. Both can be slow or shallow on very large ranges. And both inherit the arithmetic caveat: when a suite AI writes a formula, the spreadsheet computes it — reliable; when it summarizes or "reads" numbers into prose, the model is doing the math — check it.
For heavier analysis, ChatGPT and Claude accept file uploads — drop in a CSV or Excel file and converse with it. ChatGPT's data analysis mode and Claude's analysis tool both do something crucial: they can write and run actual code (invisibly to you) to compute answers, which makes their arithmetic far more trustworthy than pure text prediction. You will see phrasing like "analyzing your file" — that is real computation happening.
The privacy gate comes first, though. A customer CSV is a database of personal data. Uploading real customer names, phone numbers, salaries, or account details to a personal-tier chat tool is exactly the breach Module 8 exists to prevent. Before uploading: strip or scramble identifier columns (names, emails, phone numbers, IDs), or work on an enterprise-tier tool your company has approved for that data class. Region-product-amount-date tables usually survive redaction with full analytical value intact.
Once the file is in, the conversation is the analysis:
You no longer need to know what a pivot table is called to do what a pivot table does. Four question shapes cover most business analysis:
| Question shape | Plain-English ask | What you get |
|---|---|---|
| Totals | "Total sales by region and by product category" | The summary table a pivot would give |
| Trends | "Monthly revenue for the last 12 months — is it growing, flat, or declining, and at what rate?" | Direction and pace, the memo's spine |
| Segments | "Compare average order value between new and returning customers" | The comparisons that explain why |
| Outliers | "Which reps or rows deviate most from the pattern, and how much?" | Where the story (or the error) hides |
When you chart the answers, one rule: one chart, one message. "North region drove all Q2 growth" is a chart; "here is all the data" is not. Ask the AI which chart type fits your point — trends want lines, comparisons want bars — and be suspicious of anything decorative it offers. A 3-D exploding pie chart has never clarified a business decision.
The analysis is not the deliverable — the memo is. Managers act on sentences, not spreadsheets. AI is excellent at drafting the narrative if you feed it verified numbers rather than asking it to remember them: "Using only these figures I've verified [paste them], draft a variance commentary: what moved, the likely driver, and the recommended action, in that order." Then apply the "so what" test to every paragraph: a reader should learn what happened, why it matters, and what you propose doing. "Sales rose 12%" is data. "Sales rose 12%, driven entirely by the North launch — extending it to West could add similar growth next quarter" is a narrative someone can act on.
Draft the narrative from your verified numbers, not from the AI's memory of the conversation. Long analysis chats drift — a figure computed twenty messages ago can resurface subtly changed. Re-pasting the final verified figures into the drafting prompt costs ten seconds and eliminates the whole failure class.
Why does a system this fluent miscount? Because a language model predicts likely text, and "1,347,209" is not a likely-text problem — it is a computation. When the model runs actual code (ChatGPT and Claude analysis modes) or writes a formula the spreadsheet executes, arithmetic is trustworthy. When it computes "in its head" — summarizing a table in prose, totaling numbers inside a chat reply, quoting a percentage in a draft memo — errors are routine and, worse, plausible-looking: the wrong total will have the right number of digits.
So build reconciliation into every analysis, every time:
The costliest spreadsheet failure is not a wrong analysis — it is a wrong number that travels: into your memo, your manager's deck, the board pack. Numbers get laundered into credibility with each forward. Your reconciliation pass is the only checkpoint the error will ever meet.
AI removes the formula barrier and the pivot-table barrier, but never the arithmetic risk: trust computed answers, distrust prose answers about numbers, and reconcile totals, spot-checks, and row counts before any figure enters a memo. You own every number you forward.