Build1 publisher3 min readPublished
Hand-check ten spreadsheet answers before the copilot sees a number you cannot verify
A sanitized workbook keeps the blank rows, duplicate IDs and text-formatted numbers, and the ten questions put to the assistant all have answers someone already computed by hand. Every formula and cell range gets saved.
The Engineer · Build desk

What happened
- A dev.to checklist for spreadsheet AI assistants prepares ten questions with hand-checked answers, runs them without showing the expected numbers, and saves the formula, query, cell range, chart data and explanation for each response.
- The named question types cover a date-range sum, a pivot by product or region, a duplicate count on a stable ID, a two-table join, a trend with a changing denominator, an ambiguous term, an expected error and a footnote answer.
- The test corpus is a small sanitized copy of a real workbook that keeps the blank rows, duplicate IDs, refunds, text-formatted numbers and ambiguous labels, with personal and confidential data removed.
- The workbook copilots are grouped together: Copilot inside Microsoft 365 Excel with Python in Excel, Google Sheets AI in Workspace, and WPS AI for China-based teams working with WPS files.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Scoring only the final total will pass an assistant that summed the wrong range, and dev.to's warning is that the same formula fails on next month's data.
- decision The piece puts the layer choice ahead of the vendor choice, so the first call is whether the workbook, a typed-table product, a SQL translator or a semantic model owns the question.
- exposure Switching on natural-language Q&A over a semantic model hands every business user a confident paragraph built on whichever of three definitions of retention the model happens to encode.
- cost A dedicated NL2SQL product has to beat the SQL a general coding assistant already writes by enough to justify another tool and another place to manage database access.
The eight named question types break in different places. The checklist supplies the list of places to look: the selected range, filters, hidden rows, date boundaries, and whether the formula copied a text value instead of calculating it [6]. Those five checks are the answer key for the date-range sum. The duplicate-count question specifies a stable ID because pointing the same count at a name column returns a different number [3].
The checklist leaves two of the ten slots open. The list names eight types [3] against an instruction to prepare ten questions [2], so those two are yours to write, for whatever is specific to your workbook [18].
Saving the formula, the cell range and the explanation is the step people skip. A demo opens on a clean table and a friendly request [22]. The rule dev.to states for the workbook layer is that the assistant writes a candidate formula and the user owns the number [6]. The piece says most spreadsheet AI failures are diagnosable if you preserve the intermediate artifact [17].
Two of the ten questions are definition tests. An ambiguous term such as "active customer" and an answer that lives in a footnote both test whether the assistant can tell a definition from a calculation [3][1]. The gap survives a better model. "Revenue" may mean invoiced revenue, paid revenue, or recognized revenue, and according to dev.to no model can infer a business definition reliably from a column named amount [13]. Power BI AI and Qlik Sense AI put natural-language questions on top of a semantic model. The piece calls that the right layer when many business users need consistent answers from governed data, and also the layer where missing governance becomes impossible to hide [14].
For the workbook products, the transfer condition is stated plainly: test the exact workbook, locale, formulas and sharing model the team uses [9]. Excel AI is not automatically better because it can call Python, and Sheets AI does not become a warehouse because it can summarize a range [9].
Below the spreadsheet, the products change shape. A shared customer list, inventory table or project ledger that has outgrown a sheet is often a data-shape problem [21]. Teable answers that with typed fields, views, permissions, APIs and AI-assisted table operations, self-hostable when records have to stay inside a controlled environment; it is not a full BI platform [10]. AI2SQL and SQLAI translate questions into SQL and can explain, optimize or repair the query, and their value rests on one prerequisite: the person running the query can read the SQL [11]. The advice at that layer is a complete schema with field comments, read-only credentials, and known-answer queries before anything else [12].
What to watch
- Whether Excel, Sheets or Power BI expose which model and which data date produced a given answer, as the checklist asks for before enabling natural-language Q&A.
- Whether teams that keep the saved formulas and ranges find their failures cluster on date boundaries or on text-formatted numbers.
- Whether a metric dictionary and named data owners land before semantic-model Q&A does, or after the first disputed number.