Coverage of AI in business tends towards transformation. The reality in small companies is narrower, duller, and considerably more useful than the framing suggests.
What gets used
Writing that nobody wants to do. Product descriptions, listing copy, routine emails, first drafts of documents that will be edited anyway. This is the largest category by a wide margin.
Sorting and extraction. Pulling structured information out of unstructured text – invoices, forms, enquiries – and putting it somewhere useful. Unglamorous, immediately valuable, and easy to verify.
Translation, particularly for businesses selling across borders where hiring a translator for every product update was never affordable.
Answering repeated questions, either internally or through a support channel where the same twenty questions account for most of the volume.
What does not get used
Anything requiring the output to be correct without checking. Small businesses discovered quickly that a confident wrong answer costs more than no answer, and they calibrated accordingly.
Anything requiring integration work. If it does not work through a browser or an existing tool, it does not get adopted, because there is nobody whose job is to make it work.
The economics now
Prices fell sharply through 2026, which moved the threshold. Tasks that were not worth automating at previous rates are worth automating at current ones, and the calculation changed for volumes that are small by any enterprise standard.
The realistic advice
Pick the single task that consumes the most time and requires the least judgement. Automate that one thing. Measure whether it actually saved time, including the checking.
Most failed adoptions come from starting with something ambitious rather than something boring. The boring one pays for itself and teaches you where the limits are.