Using AI on Excel and Spreadsheets

Using AI on spreadsheets is not making charts pretty. It is locking column cleanup and classification rules, then taking formula and pivot drafts. Ask an agent for schema work, dedupe, and a pivot draft. A person still closes totals, row counts, and spot checks. This article does not list prices for paid sheet-AI tools.

Replace “analyze this table” in the search box with the S04 skeleton: goal, scope, done-check — applied to a sheet.

How do you clean and classify?

One-line answer: Lock column names, types, and empty-cell rules first. Classification labels may use only a human-approved list.

Inputs a person fills before the agent runs:

  • Source: CSV/workbook name, sheet tab, header row.
  • Schema: meaning and allowed values per column. Dates need a format; money needs currency and decimals.
  • Banned: overwriting the source, filling blanks by guess, exporting PII columns.
  • Artifact: a cleaned table (or new tab) plus a change log of what changed.

Sample instruction: “Read only sales_raw. Write a new sales_clean tab. Keep date, region, sku, qty, amount. region must be from {Seoul, Gyeonggi, Other}; map everything else to Other. Do not fill empty qty — mark those rows [needs check]. Do not edit the source tab.”

A useful cleanup artifact:

  1. One header row; empty/full-duplicate rules are written down.
  2. Classification columns follow a closed list (or mapping table).
  3. Dropped or merged columns appear in the change log.

If the agent invents “nicer” categories, put the list back. Classification is a business decision, not model taste.

What about formulas and pivots?

One-line answer: Write the aggregation questions and row/column/value fields first. Take formulas and pivots as drafts, then cross-check the same question two ways.

“Analyze this” only grows prose. Split the work:

  • Question session: at most three observable asks (“monthly revenue by region”, “top 10 SKUs by qty”).
  • Formula session: candidate SUMIFS/XLOOKUP with arguments as a table. Prefer column names over fragile cell addresses.
  • Pivot session: a pivot draft with row, column, value, and filter fields named. Charts wait until the pivot is locked.

Sample done-check: “For three questions, the pivot tab and a SUMIFS check column agree. Changing filters (period, region) moves both paths together. No guessed numbers.”

Search-shaped asks (“give me insights”) produce sentences. Job-shaped asks lock fields, filters, and the cross-check. Even if an agent builds the pivot, a person confirms whether the value field is sum or count.

If cleanup and pivot agents run in parallel, they must read the same sales_clean schema. If each renames columns, you cannot merge.

How do you verify?

One-line answer: Cross-check row counts, totals, and a few sample rows against the source. Do not accept unexplained outliers.

Human gates:

  • Row count: clean rows equal source rows kept by the rules; mismatches must match the change log.
  • Totals: amount/qty grand totals match the source with no filter — or excluded rows are logged.
  • Samples: top, bottom, and five random rows by eye — date parse errors, comma/decimal issues, full-width digits.
  • Pivot vs formula: if answers diverge, distrust the pivot first and inspect filters and the value field.
  • Export: identifiers still in a summary; shared summaries must drop those columns.

Plan prices for sheet-AI products move on vendor pages, so they are not frozen here. Prefer tools that let you review formulas/pivots as text and keep the source read-only.

Minimum path: (1) agree schema and classification lists, (2) build a clean tab, (3) draft pivots/formulas for the question list, (4) row count, totals, samples. If the agent paints a dashboard, do not skip 4.

Sheets behave like code. The schema is the diff; charts are the skin. Delegating skin first keeps the search-only habit.

Sources

  • S03 role split and S04 job-shaped instructions applied to a spreadsheet artifact.
  • Paid sheet-AI pricing: vendor pages (not pinned in this article).