Analyzing Data with AI, and Checking the Math Yourself

Ask a chat assistant to add 47 numbers in its reply and it may well get the answer right. It may also drop a digit, because a language model writing a sum is predicting what a sum looks like. Ask the same tool to load your spreadsheet and run code, and the arithmetic becomes exact. The code can still add the wrong column. So data work with AI holds up only when the machine computes, shows the method, and a person checks the result.

Tools that run code

As of October 2026, the major assistants can analyze uploaded spreadsheets by writing and running code. OpenAI's help center says ChatGPT writes and runs Python in a notebook environment, accepts .xls, .xlsx and .csv files among others, cannot reach the web from inside that environment, and recommends that users review the generated code, outputs and assumptions. Claude accepts CSV and XLSX uploads up to 500 MB per file in chat and 30 MB in projects (XLSX requires the code execution and file creation setting to be on). Google announced in July 2026 that Gemini Notebook would get code execution for data analysis, rolling out by plan. Limits change; each vendor's help page lists the current ones for each plan.

The visible code is the point. When the tool shows you the steps, you can see which column it summed, which rows it filtered out, and whether it divided by the right thing. People who don't read code can ask the tool to explain each step in plain words, numbered. The check is on the recipe, since the cooking is exact.

What a show-your-work request looks like

[complaints_2025.csv attached]

This file lists code-enforcement complaints for the invented city
of Fenwick in 2025. Before analyzing:
1. Tell me the number of rows, the column names, and the date range.
2. Count exact duplicate rows and rows with the same address and
   date. Do not remove anything yet; report the counts.
Then, after I confirm:
3. Count complaints by neighborhood, with duplicates removed.
4. Using the household counts I'll paste below, give complaints per
   1,000 households.
Show each step's code and explain it in one sentence. Give final
numbers in a table I can copy.

Splitting the job shows the raw data's shape before any conclusions. If the row count doesn't match what the city said it released, that mismatch matters more than any result that follows.

Worked example: counts versus rates

Fenwick (invented) released 1,412 complaints. The AI found 37 exact duplicates and 61 same-address, same-day pairs that looked like one complaint entered twice. Lena Arvo, a neighborhood association volunteer, called the code office, which confirmed that its system logs a phone complaint and its follow-up as separate rows. After removing the 37 exact duplicates and the 61 double entries, 1,314 remained (1,412 minus 37 minus 61).

Lena's hand check: 508 + 372 + 291 + 143 = 1,314. Good. Riverside: 372 divided by 2,480 = 0.15, times 1,000 = 150. Northgate: 508 divided by 6,350 = 0.08, times 1,000 = 80. Citywide: 1,314 divided by 15,570 = 0.0844, or 84.4 per 1,000.

Northgate has the most complaints. Riverside has nearly twice the rate. Those are two different findings, and only the second one says something about the neighborhood instead of its size. The AI's first summary opened with "Northgate tops the complaint list," which was true and mostly useless.

Then she asked one more question no spreadsheet could answer: who files the complaints? The code office said one Riverside landlord dispute produced about 90 complaints from two callers. That's 90 of Riverside's 372, almost a quarter. Data analysis gets you to the question. A phone call answers it.

The errors to hunt

Charts

AI tools draft charts quickly. A chart deserves the same hard reading as text, whether it comes from an AI or from an agency slide deck.

• Bars start at zero. A bar chart with an axis starting at $3 million makes a 5 percent difference look like a tripling.
• Labels and source line. Every chart says what's measured, the unit, the years and where the data came from.
• Rates for unequal groups. Neighborhoods of different sizes get per-household bars, not raw counts.
• One message. A chart whose point can't be said in one sentence shows too much.
• Re-plot one point. One bar's height checked against the table catches a mislabeled chart.