How to analyze spreadsheets with AI in Whizi

Profile the data before you ask it anything

The most common way spreadsheet analysis goes wrong has nothing to do with AI. It is that the file contains 14 rows with a trailing space in the region name, two date formats, a subtotal row somebody left in the middle, and 300 blanks in the column you are about to average. Ask your question first and you will get a confident answer computed over garbage.

So the first prompt is the same on every file.

Prompt: the profile

Profile this file before we analyse anything. Return: row count, column names with inferred type, the count and percentage of missing values per column, the number of exact duplicate rows, the distinct values for every column with fewer than 25 distinct values, the min and max of every numeric and date column, and any column where the values look inconsistent (mixed formats, trailing whitespace, mixed units, mixed date formats). Do not analyse or interpret yet. Just tell me what is in the file and what looks wrong with it.

That single prompt catches most of what would otherwise ruin the analysis. The distinct-values list in particular is where you discover that "UK", "U.K.", and "United Kingdom" are three separate regions in your data.

Prompt: fix what it found

Standardise the issues you identified: trim whitespace, unify [column] to a single format, and consolidate these variants: [list them]. Show me the mapping you applied as a table before applying it. Do not drop any rows without telling me which and why.

Ask questions that produce checkable answers

Vague questions get vague answers. The prompts that work name the column, the operation, and the output format, and they ask for the working.

Prompt: aggregation with the working shown

Group by [column] and calculate [metric]. Return a Markdown table with the group, the count of rows in it, and the metric. Below the table, state exactly how you computed the metric, which rows you excluded and why, and how you handled blanks. Sort descending by [column].

Prompt: the cohort question

For each [cohort dimension, for example signup month], calculate [metric] at [interval]. Show the cohort size alongside every figure. Flag any cohort with fewer than [n] rows as too small to interpret rather than reporting a percentage for it.

That last instruction prevents the single most misleading output in spreadsheet analysis: a cohort of four users reported as "75% retention" and sitting in a table next to a cohort of nine thousand.

Prompt: anomaly hunting

What is suspicious about this data? Look for: values outside a plausible range, sudden discontinuities in a time series, columns where the distribution changes partway through, rows that are exact or near duplicates, values that appear too round, and anything that suggests a change in how the data was collected. For each, quote the specific rows.

That is the highest-value prompt on this page and it has no equivalent in a normal spreadsheet workflow. It routinely finds the day the tracking broke, the vendor who started reporting in a different currency, and the duplicate import that inflated a quarter.

Prompt: the chart specification

Recommend the right chart for this question and this data shape, and explain why the obvious alternative is worse here. Then give me the specification: chart type, x, y, series, aggregation, sort order, and axis treatment. Do not use a dual axis. Do not truncate a bar chart axis.

Where the arithmetic actually goes wrong

This deserves a straight answer, because "AI is bad at math" is both true and unhelpfully vague.

A language model reasoning over numbers in text is doing pattern completion, not calculation. It is reliable at describing structure, categorising rows, and identifying which computation you need. It is much less reliable at performing a long chain of arithmetic across hundreds of rows, and it fails silently: the output is a well formatted table of wrong numbers with no distress signal.

The practical rules:

• Ask for the method, not just the result. "State exactly how you computed this and what you excluded" makes the error visible when there is one.
• Spot check one group by hand. Pick the smallest group in the output table and verify it in the actual spreadsheet. If it matches, the method is probably right; if it does not, nothing in the table can be trusted.
• Cross-check anything consequential with a second model. Two independent models arriving at the same number is meaningfully better evidence than one model repeating itself. Whizi's side-by-side view exists for exactly this.
• Watch for double counting. Subtotal rows left in the file and one-to-many joins are the two usual culprits, and the model will not know they are wrong.
• For anything that must be exact, ask for the formula or the code. Give me the Excel formula or give me the pandas code puts the arithmetic in a deterministic engine, and you keep the model for the part it is good at, which is knowing which computation to run.

That last one is the real answer for financial or reporting work. Use the model to design the analysis, use your spreadsheet or a script to execute it.

Which model reads which file, and how big is too big?

CSV and Excel both upload directly, and the three models split the work cleanly.

Model	Context window	Credits per message	Use it for
GPT-5.6 Terra	1M tokens	4	Strict formats: clean tables, JSON, exact column mappings
Claude Sonnet 5	1M tokens	10	The cross-check pass on multi-step aggregation
Gemini 3.5 Flash	1M tokens, about 1,900 manuscript pages	8	The largest files, or a spreadsheet plus a PDF in one thread

Context and credit figures come from Whizi's model cost index, list rates fetched 2026-08-20.

A few practical notes:

• Give it a clean rectangle. One header row, no merged cells, no blank spacer rows, no notes in column K. Multi-header formatted reports confuse extraction more than any file size limit.
• Send the raw sheet, not the presentation sheet. The version with the formatting and the subtotals is the one that produces double counting.
• Very wide files benefit from a column list first. Ask what each column means before analysing if the names are cryptic, and tell the model what it got wrong.
• For very large files, use Gemini 3.5 Flash, which reads the same 1M token context as Claude Sonnet 5 and GPT-5.6 Terra at the lowest per-answer cost of the three. Beyond that, sample deliberately: analyse a random 5000 row sample and tell me the sampling error I should expect on each figure is better than silently truncating.
• Strip personal data first. Names, emails, and identifiers are almost never needed for the analysis, and removing them is faster than arguing about whether you were allowed to upload them.

The mechanics of uploads are covered in uploading documents.

The full eight-step sequence on a real file

The whole workflow on a real file, in order:

1. Upload. Run the profile prompt. Read the distinct values and missing counts carefully.
2. Fix what it found, reviewing the mapping table before it is applied.
3. Ask for a summary of what the data is about and the three questions it thinks you should be asking. This step is quick and often reframes the analysis.
4. Ask your actual question, with the method and exclusions required in the answer.
5. Run the anomaly prompt, always. This is where the surprises are.
6. Spot check the smallest group by hand.
7. Cross-check the headline number with a second model.
8. Ask for the chart specification, or the formula if the number needs to be exact and reproducible.

Steps 1, 5, and 6 are the ones people skip, and they are the ones that separate an analysis you can defend from a nicely formatted guess.