Answer
How do you use AI with a spreadsheet?
Have it write the formula or transform a copy. Do not have it read numbers out and total them in prose.
Use it to write formulas, clean and restructure data, and explain what an inherited sheet does. Do not use it to read figures out of a sheet and calculate in prose, because that arithmetic is unchecked.
The reliable division is between generating a mechanism and performing a calculation. A formula is a mechanism: it is visible in the cell, it recomputes, and if it is wrong the error is systematic and usually obvious the moment the numbers look odd. A figure produced in prose is an assertion, and there is nothing to inspect. The same system that writes an excellent formula will confidently misadd a column, and the difference is not capability, it is whether the arithmetic was executed or described.
The strongest use is therefore formula and structure work. Describe what you want computed in plain words and have it produce the expression; ask what an inherited formula does; ask why a nested condition returns the wrong result on one row. This is a translation task between description and syntax, which is exactly the shape these systems are strong at, and the output is verifiable by running it.
The second strong use is cleaning. Inconsistent name formats, addresses in one field that should be four, dates in three conventions, categories that mean the same thing spelled five ways. This work is slow by hand, mechanical to describe, and inspectable afterwards by sorting the result and looking. Do it on a copy, keep the original column, and compare, which turns a trust question into a diff.
The third is explanation. A spreadsheet inherited from someone who has left is a genuine business problem, and having the logic described in prose — what feeds what, which cells are inputs, where a hard-coded number is hiding — is real value with a low failure cost, because you will verify the account against the sheet as you read it.
The failure everyone hits is the summarising question: what does this data show, which region is worst, what is the trend. These invite the system to read values and reason over them in text, and the reasoning is not arithmetic. Where the question is genuinely analytical, the safe form is to ask for the calculation as a formula or a query, run it, and read the result off the sheet. That takes a minute longer and produces a number that has actually been computed.
One structural point about size. A sheet supplied as text into a conversation is a snapshot, and a large one will not fit; what fits is whatever fitted, and there is usually no announcement about the remainder. If a job depends on all the rows, the job needs a tool that queries the file rather than a paste, and this is one of the few cases where an integration is clearly worth its permission cost.
A formula can be inspected, re-run and disagreed with; a number asserted in a sentence has none of those properties and looks exactly the same.
Siddharth Sharma, Context Theory
Related questions
Are the built-in AI features in spreadsheet software different?
They are better positioned, because a feature that writes into a cell produces a formula that recomputes, and one that fills a column can be checked against neighbours. The same caution applies to anything that returns a natural-language summary of a range: whether the number was computed or composed is the question, and the interface rarely tells you which.
How do you check a formula it produced?
Run it against rows where you know the answer, including the awkward ones: an empty cell, a zero, a duplicate, the row at the boundary of the range. Most generated formulas are correct in the ordinary case and wrong at an edge, which is also true of hand-written ones. The edge rows are the test, and they take a minute.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Odds of qualifying a lead — replying within the first hour vs after it | 7× | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Software | A generated formula is a mechanism that can be inspected and re-run while a figure asserted in prose is an unexecuted claim, so the reliable division in spreadsheet work is between producing a calculation and performing one, not between easy and hard questions. | Asking the same system to total a pasted column in prose and to write the equivalent formula, then comparing both against the sheet. |
| Workflow | Data cleaning is safely delegable because the result is inspectable by comparison: performing it into a new column beside the retained original converts a trust question into a difference that can be sorted and read. | Running a cleaning pass into a parallel column and sorting on the pair to surface disagreements. |
| Response | A spreadsheet supplied as pasted text is truncated silently when it exceeds what fits, so any job that depends on every row requires a tool that queries the file rather than a paste, regardless of how the truncation appears. | Pasting a long sheet and asking for the row count, then comparing it against the file. |
| Buying behaviour | Generated formulas are typically correct in the ordinary case and wrong at boundaries, so the efficient check is a small set of known-answer rows containing an empty cell, a zero, a duplicate and the range edge rather than a general review. | Testing any generated formula against those four row types before applying it. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one