Spreadsheets run a surprising amount of small-business life - the pricing, the budgets, the stock counts, the figures that decisions are actually made on - and writing the formulas that make them work has always been a fiddly, error-prone chore. Getting a calculation to reference the right cells, handle the edge cases and produce the number you meant takes patience and a tolerance for cryptic error messages. So it is genuinely useful that AI assistants can now write these formulas from a plain description: say what you want calculated, and get a working formula back in seconds, without wrestling with the syntax. For anyone who finds spreadsheets more obstacle than tool, this removes real friction.
The danger is specific to the medium, and it is worth naming clearly: a spreadsheet formula can be completely wrong and show no sign of it. Unlike a sentence, which reads as awkward when it is off, a formula that calculates the wrong thing usually still runs, produces a plausible-looking number, and sits in the cell looking exactly as authoritative as a correct one. There is no error message, no red flag - just a number that happens to be wrong. If the formula references the wrong range, mishandles a case the description did not mention, or interprets the request slightly differently from what was intended, the result is not a crash but a quietly incorrect figure that then flows into every total and decision built on top of it.
This matters more in a spreadsheet than in most other AI outputs because of what spreadsheets are used for and how their errors propagate. A wrong figure in a cell is rarely the end of the chain; it feeds a total, which feeds a summary, which informs a price, a budget, a decision. A single mis-written formula can therefore corrupt an entire model silently, and because everything downstream still computes and still looks normal, the error can persist for a long time before anyone notices - if they notice at all. The very fluency of the AI, producing a clean formula instantly, makes it tempting to accept the result without the scrutiny that a hand-written formula, produced slowly and with visible effort, naturally invited.
The discipline that keeps this useful rather than hazardous is to verify the formula against something you can check independently, in proportion to what the number is used for. For a throwaway calculation, glancing at whether the result looks roughly right is enough. For a formula that will sit at the heart of a pricing model or a budget, it is worth testing it on a small case where you already know the answer - feeding in figures you can compute by hand and confirming the formula returns what it should - before trusting it across the real data. The AI is an excellent and fast writer of formulas; it is not a guarantor that the formula computes what you meant. Treating it as a quick draftsman whose work you spot-check on a known case, rather than as an oracle, is what lets a business capture the speed without building its important numbers on a formula that runs perfectly and calculates the wrong thing.