Can a local AI read spreadsheets?
Yes, with a caveat worth knowing before you hand over the household budget or the lab's measurement log.
Yes, with a caveat worth knowing. A CSV opens as plain text and works well. An Excel workbook needs its cells extracted, usually into text or a table, and the model reads that conversion rather than the grid itself. Numbers survive the trip; visual layout does not, and merged cells can confuse the reading order. For heavy spreadsheet work, ask targeted questions of specific columns rather than expecting the model to audit the whole sheet.
Text in a grid.
A spreadsheet is text arranged in a grid, and a model reads text, not grids. The two common cases differ in how much work the arrangement takes to undo. CSV is already linear text, one row after another, so it passes through with almost no conversion, and it sits alongside the formats covered on what a local AI can read. Excel files are containers: a workbook holds sheets of cells in a compressed structure, so a local tool parses the file and turns the cells into something textual before the model sees any of it. Both conversions run on your machine, with no upload, because that is what every format in a local loop has in common.
What survives the conversion.
The values usually arrive intact. The presentation rarely does: which cells were merged, how columns were sized, what color flagged what, all of that lives in the grid and is lost or flattened when the sheet becomes text. Reading order can bend around merged cells and multi-row headers, and a wide sheet can simply be longer than what fits in the model's working memory at once, which is why retrieval decides which rows reach the model rather than the whole sheet arriving at every question. None of this is a local-versus-cloud difference; a cloud tool faces the same conversion and the same limits, plus the upload.
How to ask about a sheet.
The reliable pattern is narrow questions. Ask what a specific column says, which rows match a condition, or what a range contains, and treat the answer's citations as pointing at the extracted rows they came from. Be slow to trust arithmetic the model performs itself: totals, averages and rates across many rows are where a language model is least dependable, so have those computed in the spreadsheet or a script where the result is checkable. A question about what the data means, once the values are quoted and verified, is where the model earns its keep, and the habit of checking claims against the source applies here as much as it does on every other kind of answer.