Transcription just got cheap and good. Four things an SME can do with that
OpenAI released new speech-to-text models yesterday. Dictation is the most underrated AI use case in a trade or field business.
On 20 March 2025 OpenAI released new audio models, including speech-to-text systems reported to improve on the previous generation in word error rate and language recognition. Transcription has been quietly getting better for two years; this is another step.
It is also the AI capability least discussed and most immediately useful to companies where people work with their hands.
Four uses that pay for themselves quickly
The site report spoken in the van. A technician finishing a job describes what they did, what was replaced, what needs a follow-up visit. Five minutes of speech becomes a structured report before they reach the next customer. The alternative is a form filled in at eight in the evening, badly, or not at all.
The customer call note. Thirty seconds of dictation after hanging up, turned into a CRM entry. This is the single most common gap in SME customer records: the call happened, nobody wrote it down.
Meeting minutes that actually get written. With one caution below.
The dialect problem, reduced. Swiss German dictation is genuinely harder than standard German, and results vary by speaker and by product. It is markedly better than two years ago and still worth testing with your own people before promising anything.
The two rules
Tell people they are being recorded, every time. Recording a conversation without the knowledge of everyone in it is not a grey area in Switzerland. For internal meetings, say it at the start. For customer calls, ask. A transcript obtained without consent is a bigger problem than the note it saved.
Decide what happens to the audio. The recording is often more sensitive than the transcript, and it is the part people forget to delete. Set a retention rule — delete the audio once the text is approved is a reasonable default — and write it into the same page as the rest of your AI rules.
Why this one is worth starting with
Because unlike most AI projects it needs no integration, no master data cleanup and no vendor negotiation. One person, one phone, one week of trying it on real jobs. If it works for them, it will work for the others.