When Kuwait calls need a transcript, not an agent

The job is a recording you already have, turned into text a person can search. Speaker, time, a line you can quote. Not a bot on the next call.

Teams ask for “AI voice.” That often means an agent that answers. This job starts after the call ended.

What we decide on a first call

Where the audio lives — a PBX export, a phone, a meeting tool. How many speakers, and whether you must name them. Arabic, English, or a mix on one file. What “good enough” means: a searchable log, or a line you will send to a customer. Who may hear the raw file. Where the text has to sit: the CRM note, a folder, a sheet.

If you do not own the recording, or people were not told it is kept, we stop. That is not a model problem.

If the files are yours and the stall is listening again, AI voice generation includes speech-to-text on the way in: a transcript, a speaker split, a clock. We will say when a noisy shop floor will fail. We will not promise a court record from a pocket recording.

This is not a phone or IVR agent. That job talks back. A transcript does not pick up. It is also not a Slack digest from systems you already run. Digests summarise rows. This job starts from audio. It is not answers grounded in a doc corpus either. Those docs were typed first.

When we tell you to keep listening

If the call is the sale and the closer will not trust a line they did not hear, keep the ear. If you only keep ten recordings a week, a person and a note is still honest.

Kuwait calls often switch Gulf Arabic and English mid-sentence. Name that on the call. Do not ask a model to “clean it up” into one language if the quote has to stay true.

AI consultancy writes the source, the speaker rule, and the destination first. A batch comes after you can fail a file for the right reason.

Bring one recording you already replay and the note a person typed after it. Book thirty minutes or write hello@jamilglobal.com.