How to Use Claude to Analyse Excel Data

Start with the decision, not the workbook

A spreadsheet can support many questions, and Claude cannot choose the right one from a filename. Write the decision in one sentence, name the audience, reporting period, currency, acceptable precision and action that may follow. Tell Claude whether you want a data-quality audit, descriptive analysis, scenario comparison or a management brief. Do not ask it to find insights without a decision boundary; that invites decorative patterns.

Add a date cutoff and source hierarchy. The workbook is the source for its observed values, while product documentation, a regulator or a supplied report may be the source for external facts. Ask Claude to label every statement OBSERVED, CALCULATED, ASSUMED, INTERPRETED or UNKNOWN. This prevents a plausible interpretation from quietly becoming a reported fact.

Prepare a safe working copy and data contract

Duplicate the workbook before uploading it. Remove passwords, unnecessary identifiers, confidential columns and hidden sheets that are not needed, or replace them with stable labels. Record the file date, owner, source system, currency, timezone and whether amounts are gross, net, tax-inclusive or tax-exclusive. If a field is a rate, state whether it is a percentage, decimal or annualised value.

Request a data contract before interpretation: one row per sheet and column, data type, unit, period, example value, expected range, allowed blanks, key field and known caveat. Ask Claude to stop when it cannot establish a definition. This exposes duplicate IDs, mixed date formats, totals mixed with detail rows and formulas pasted as values. Keep the dictionary beside the workbook so another reviewer can follow the analysis.

Inventory sheets, formulas and missingness first

Your first prompt should not ask for a recommendation. Ask Claude to list every sheet it can read, its apparent purpose, headers, formula regions, blank patterns, duplicate keys, date coverage and suspicious totals. Require it to distinguish what it directly inspected from what it inferred. If the file is too large or a sheet did not parse, that is a result, not an invitation to guess.

Ask for a validation sample: rows from the beginning, middle and end, rows with blanks, unusual values and rows that feed a total. Compare the response with the visible workbook. Anthropic's current documentation says XLSX support depends on the analysis tool being enabled and that non-PDF documents are text-extracted, so layout-heavy interpretation needs extra caution.

Use Claude for formula explanation and cleaning plans

Claude can explain a formula in plain language, find repeated formula patterns and draft a cleaning plan. Ask it to return the sheet, cell or column, original formula, plain-English meaning, edge cases and a safer test. Do not ask it to silently rewrite the original workbook. For cleaning, require a transformation table with input example, output example, rule, rows affected and whether the change is reversible.

Common checks include trimming whitespace, standardising labels, parsing dates, identifying numbers stored as text and distinguishing zeroes from blanks. Ask Claude to preserve raw values and write cleaned output to a new sheet or file. Test each rule on known rows, then compare row counts, totals, unique keys and null counts before and after. A cleaner-looking sheet is not automatically a more accurate one.

Make calculations reproducible

For every reported number, ask for a calculation ledger: input sheet and range, filter, aggregation, unit conversion, rounding and output. If Claude supplies a formula or code, place it in a controlled copy and run it against known values. Ask for an independent cross-check, such as a pivot total versus a filtered SUM or a manual sample versus a formula result. Differences should stop the workflow until explained.

Keep observed data, assumptions and scenarios on separate sheets. If you ask for a forecast, state that it is a scenario and list assumptions rather than allowing history to masquerade as a prediction. Show which variable changes each result. Claude can structure the comparison; you own the arithmetic and decision.

Separate findings, interpretations and recommendations

A finding might be that returns were 12% of orders in the supplied period; an interpretation might be that returns were concentrated in two SKUs; a recommendation might be to review those descriptions. Each layer needs a different evidence standard. Require a source location beside each finding and a caveat beside each interpretation.

Choose charts for the question, not because the tool can make them. A time series needs consistent periods, a comparison needs a meaningful baseline and a relationship plot does not prove causation. Ask for a chart specification with title, axes, units, filters, exclusions and the decision it supports. Render it in your spreadsheet, then inspect labels, outliers, zero baselines and truncated axes.

Use staged prompts and explicit checkpoints

A reliable session has checkpoints. Stage one establishes the contract and inventory. Stage two audits quality and formulas. Stage three performs the requested calculation. Stage four drafts the narrative. Stage five runs an adversarial review. At each checkpoint ask Claude to report sheets used, assumptions, unresolved questions and what it could not inspect. Save the approved prompt and output with the workbook version.

If you have several workbooks, give them roles and ask Claude to identify join keys before combining them. Require a row-count and duplicate-key report after every join. Never let a fuzzy name match silently become a confirmed relationship. If the join is uncertain, produce a review list for a human.

Protect confidential spreadsheet data

Spreadsheets often contain payroll, customer, health, pricing or contract information. Use the minimum columns needed. Replace names and IDs with stable pseudonyms, remove credentials and keep free-text notes out unless essential. Confirm your organisation's approved account, retention and sharing rules before uploading. A product limit is not an organisational permission.

If unsure whether a field is sensitive, treat it as sensitive. Use an anonymised sample for prompt design, then move the validated method into an approved environment. Do not paste a link that grants broader access than the task needs. Keep a record of the reviewer and final storage location.

Review the workbook and report separately

A chat answer can be correct while the workbook remains wrong, and a workbook can be correct while the report overstates what it shows. In the workbook, inspect formulas, filters, hidden rows, named ranges, chart sources, dates, units, totals and conditional formatting. Open it in a second viewer if possible. In the report, check every material number against the ledger and every current claim against its primary source.

Ask Claude for an omission and contradiction pass, then perform your own approval. Confirm the original is preserved, the output filename identifies the version and unsupported statements are marked. For finance, safety, employment, health or legal decisions, add qualified review.

Finish with an evidence and handoff pack

A useful analysis should leave behind more than a paragraph. Keep the source copy, cleaned copy, data dictionary, calculation ledger, assumptions, chart specifications, source register, change log and approval note together. State what was analysed, what was unavailable, what assumptions were accepted and what the reader should verify next.

Ask Claude to produce a short handoff for someone who did not see the chat. Include the question, method, key findings, limitations, unresolved items and exact next action. This makes the work inspectable and reduces the chance that a future reader treats a scenario, inference or stale number as an established fact.

Handle joins and reconciliation explicitly

Most spreadsheet errors happen at the boundaries between tables. Before asking Claude to combine orders with customers, costs with revenue or targets with actuals, define the expected grain of each table and the join key. Ask for a pre-join report showing row counts, unique keys, null keys, duplicate keys and unmatched records on both sides. A one-to-many join can multiply revenue silently; a misspelled key can turn a real customer into an unknown.

After the join, request the same counts again and reconcile totals against the source sheets. Keep unmatched rows in a separate review table instead of dropping them. If names are used as keys, treat the result as provisional unless a controlled mapping exists. Claude can suggest matching rules, but a human should approve ambiguous matches and document why. This is especially important when the output feeds compensation, credit, billing or performance decisions.

Design a repeatable monthly workflow

If the analysis repeats, turn the successful session into a checklist rather than relying on chat memory. Define the input files, expected filenames, reporting period, column contract, validation checks, output tabs, reviewer and deadline. Ask Claude to compare the current workbook with the prior period and report new sheets, changed headers, missing periods, formula drift, row-count changes and unusual movements before drafting commentary.

Keep a change log between periods. A change in a source system, category mapping or currency can explain a large variance better than a business event. Require the model to label a change as observed or hypothesised and to point to the evidence. Automate only the mechanical checks you understand. A recurring workflow still needs a named owner who can stop publication when an input is incomplete or a definition changed.

Write an executive brief without hiding uncertainty

An executive reader needs a short answer, but short does not mean absolute. Ask Claude for a one-page brief with the decision, three verified findings, two implications, one recommendation, material caveats and the next owner. Put the reporting period and units in the heading. Link each number to a calculation ledger entry or workbook range so a reviewer can trace it quickly.

Avoid phrases such as “the data proves” when the workbook is observational. Use “in the supplied period,” “is consistent with” or “requires confirmation” when those are more accurate. Put important exclusions beside the finding, not in a footnote nobody reads. Ask for a version written for a sceptical reader who will challenge the denominator, time period and comparison group. Clarity about uncertainty increases trust; it does not weaken the analysis.

Check units, dates and denominators

Before accepting a trend, ask Claude to repeat the denominator, time window and unit for every headline metric. Revenue per customer, revenue per order and revenue per active customer can all be correct numbers with different meanings. A percentage may be a share of rows, a share of value or a change from a prior period. Ask for a metric dictionary that states numerator, denominator, exclusions and rounding.

Check date boundaries explicitly. A month labelled September may contain a different timezone, posting date or service period from the comparison month. Ask the model to identify partial periods, late-arriving records, refunds, cancellations and duplicated snapshots. Keep currency conversions and inflation adjustments visible rather than burying them in prose. These checks take little time and prevent the most persuasive spreadsheet mistake: a precisely calculated answer to the wrong question.

Before handoff, ask a second reviewer to reproduce one headline number from the stated range. If they cannot do it quickly, the definition or workbook structure is not clear enough yet. Record the correction rather than silently changing the narrative.