What generative AI does: recorded calls, made searchable

What generative AI actually does with a recorded conversation is turn it into text you can search. That is three separate steps rather than one. Transcribe the audio into words, index the words so they can be found, and retrieve the right passage when somebody asks a question. Each step fails in its own way.

Somebody has just done this to South African talk radio, which makes the pipeline easy to look at.

What generative AI was pointed at

TechCentral reports that South African talk radio is now searchable, and that Shaun Trennery built the platform without writing a line of code himself.

The no-code detail is the part being covered, and it is the least interesting thing about it. What matters to an operator is the shape of the problem, which is one you already have. A large volume of spoken material holds real information, and none of it can be found by anybody.

Half of what a South African business knows was said out loud. It was said on a call to a customer, at a site visit, in a handover between shifts, or to a supplier who agreed to something on the phone. None of it is in a system. When it matters six months later, the answer is whoever was on the call, if they still work there.

Step one: transcribe

Speech becomes text. This is the step that has genuinely changed, and it is now cheap enough to run across everything rather than on selected calls.

It is also the step that fails locally, and the failures are specific.

• South African English across the country's many accents, on a line that drops.
• Code-switching mid-sentence between English and isiZulu or Afrikaans.
• Industry vocabulary no general model has seen.
• Proper nouns like Ekurhuleni, Boksburg or a customer's trading name.

A model that transcribes a studio recording perfectly will produce mush from a call taken on a handset in a warehouse.

Test it on your worst audio, not your best. A vendor demonstration on clean American English tells you nothing about a Tuesday afternoon on your order line.

Step two: index

Text becomes findable. This is the step almost everybody skips, and it is the reason most transcription projects end with a folder nobody opens.

A transcript sitting in a file is not a record. It becomes one when it is attached to the thing it is about: this call belongs to this customer, this job number, this quote, this incident, on this date, with these people on it. Without that attachment you have converted an unsearchable audio archive into an unsearchable text archive, at some cost.

Indexing is also where speaker separation matters. A transcript that does not distinguish your employee from the customer cannot answer the only question anybody ever asks about a call, which is who agreed to what.

Step three: retrieve

Somebody asks a question and gets an answer with the source attached.

The retrieval step is where a model earns its place, because the question is asked in ordinary language and the answer lives across many transcripts. It is also where the answer has to arrive with a citation: the date, the call, the timestamp. An answer with no source is a guess in a confident voice, and in a dispute it is worthless.

Insist on this in any demonstration. A system that will show you which call a claim came from is usable. One that summarises without pointing at the source cannot be checked, which means it cannot be relied on.

What it costs to run across everything

Volume is what changes the arithmetic, and cost per unit falling does not mean the bill falls.

htxt.africa reports Global Kinetic's Martin Dippenaar citing a Gartner projection on exactly this. Agentic models can require between five and thirty times more tokens per task than a standard chatbot. Call centres are among the most exposed sectors, because they handle millions of interactions a week.

Dippenaar puts it plainly. The price per token is down while the number of tokens per serious task has exploded, so the true cost per useful answer is often flat or higher than it was eighteen months ago.

A business taking 200 calls a day is not a call centre, but the same shape applies. Transcribing everything is affordable. Running an expensive model over everything, repeatedly, is what produces a bill nobody forecast.

Which stage to build first

Not the clever one.

If your calls are not recorded, or are recorded on a system nobody can export from, no amount of model quality helps. If your recordings are not tied to a customer or job number, transcription produces text with nothing to attach it to. The order of work is audio you can get at, then attachment to a record, then transcription, then retrieval.

That order is also the order of cost, from nearly free to genuinely expensive, which is a useful coincidence. Most of the value in this pipeline is unglamorous plumbing between a phone system, a customer record and a job number. Zorah does that as a staged project rather than as a purchase.

What to do on Monday

Pick one dispute from the last six months where somebody said "but they agreed to it on the phone".

Try to find that call. Time yourself. If you cannot get to the audio in ten minutes, the problem is not what generative AI does and never was. The recording is not attached to the customer record it belongs to. Businesses taking phone orders across a counter, like auto parts distribution, should start by checking whether the order line is recorded at all.