Voice agent: the thirty-second test

There is a simple test that decides almost all of these projects, and it fits in one question: does your typical call resolve in thirty seconds with three pieces of information? Booking an appointment, confirming an address, saying whether an item is in stock — yes. Understanding a fault, settling a dispute, negotiating a price — no, and no amount of tuning changes that.

The second test is technical and runs on your own recordings. Voice is markedly harder than text, and the difficulty is not in the understanding: it is upstream, in transcription, which degrades on a noisy line, on a call from a vehicle, or on a sentence that switches from darja to French halfway through.

This article describes both tests, what they reveal, and why the honest answer is sometimes not to do the project. It is an article that opens by explaining how to do without us.

Once you have the measurement, look at its shape before concluding. If the missed calls concentrate in two specific hours and the rest of the day is quiet, your problem is a staffing problem in those two hours, and one more person on that slot costs less than a project and works from next week.

If instead the missed calls spread across the whole day, over breaks, over Fridays and over holiday periods, then you are dealing with a structural absence of cover, and that is the case the agent handles well.

We run that count at scoping and hand it over whatever it shows. Sometimes it concludes in favour of human cover, and we say so — a less profitable answer for us and a faster one for you.

There is a third shape the count can take and it is more common than people think: missed calls concentrated in a season. A trade whose demand triples for six weeks gains nothing from a permanent installation; it gains from reinforced cover across those six weeks. The distinction shows immediately in the measurement and does not show at all in a sales conversation.

In writing, a darja sentence in Latin script stays readable and can be handled with synonyms. Spoken, the switching is harder: the speaker moves from darja to French mid-clause, says the numbers in one language and the street names in another, and marks no boundary between them.

Recognition systems are trained mostly on languages where that switching does not occur. They handle it by picking one language and treating the rest as noise, which produces transcripts that are plausible and wrong — the worst of both, because a visibly broken transcript would trigger a handover and a plausible one does not.

There is no general solution to this, only a measurement. The test on your recordings makes it visible; what you do afterwards is a scope decision — restrict the agent to short calls and expected words, or drop it.

There is a partial mitigation and it should be presented honestly: restricting the expected vocabulary. An agent that knows it will hear a commune name, a day of the week or a product reference recognises far better inside that narrow frame than in open conversation. It works, and it requires giving up the calls that fall outside the frame — which returns you to the thirty-second test.

A voice handover has to pass three things to whoever picks up: the reason, what has already been said, and the number. Without the number, a call that drops during the transfer is a lost customer with no way to call back, and drops during transfers are common.

Handover thresholds are set low, and lower than you would first want: on explicit request, on prolonged silence, on a second misunderstanding, and on any sign of irritation. An agent that presses once too often turns an annoyance into a bad memory.

The most useful threshold is also the simplest to describe: if the caller repeats information they have already given, the agent transfers. It is the most reliable sign that something was not understood, it requires no emotion detection, and it catches most cases before they turn unpleasant.

We count your missed calls, we test transcription on your recordings, we write the script and every one of its exits, and we go live on an overflow line. The transcription test is at our cost and it precedes the quote.

The rate we will not publish is the recognition error rate. We measure one at every client, it is accurate, and publishing it would be a category error rather than an excess of optimism: a rate measured elsewhere was measured on other lines, other handsets, other speakers and another proportion of darja, and those four variables weigh more than the engine chosen. A number on a sales page would be true somewhere and unrelated to your installation. The few figures published in this field are moreover either supplied by vendors and measured on read speech in a studio, or drawn from academic work on corpora that are not Algerian telephone audio; we have found no dated source describing this case, and we will not manufacture one.

The counterpart is that the test is free and can cost us the project. We run it before the quote, we hand you the rate per call type even when it is poor, and when it is poor we advise against the project and propose the written message fallback described above. We would rather give you a recommendation you can act on without us than a rollout that frustrates your customers for six months before being switched off.