The debate about artificial intelligence has settled into a tedious equilibrium. Executives are told either that AI will displace jobs on an apocalyptic scale or that it will finally deliver the operational efficiency they were promised. Sometimes they are told both. Each camp shares an unexamined premise: that enterprise demand for intelligence – for problem-solving capacity – is fixed.
History suggests otherwise. When the cost of a vital resource collapses, demand rarely stays flat. It rises, often dramatically. AI is collapsing the unit cost of processing information, and organisations’ appetite for intelligence is rising with it.
In 1865, the English economist William Stanley Jevons published The Coal Question, which made a counterintuitive observation. Improvements to the steam engine, most famously James Watt’s, had sharply cut the coal needed to do a given amount of work: a fully developed Watt engine used about a quarter of the fuel of the Newcomen designs before it. Intuition suggested Britain would burn less coal.
The opposite happened. Cheaper work made steam power viable in thousands of new applications, from textile mills to deep-shaft mining and railways. Coal consumption soared.
This became known as the Jevons paradox, and it is the economic engine behind AI today. AI is not eliminating enterprise demand for intelligence; it is collapsing its unit cost. When applying intelligence to a problem gets cheaper, organisations don’t do it less. They turn it on a vast reservoir of latent demand – problems that were always there but too expensive to solve.
The coyote hasn’t run off the cliff
A decade ago, radiology looked like a prime candidate for AI-driven extinction. Image-recognition models could read thousands of scans without tiring. In 2016, Geoffrey Hinton, who would go on to win a Nobel Prize, said: “I think if you work as a radiologist, you are like the coyote that’s already over the edge of the cliff but hasn’t yet looked down. People should stop training radiologists now. It’s just completely obvious within five years deep learning is going to do better than radiologists.”
Sound familiar?
It didn’t happen. Radiology has become medicine’s biggest target for AI – about three-quarters of the more than a thousand AI applications approved by the US Food and Drug Administration for medical use are in radiology – yet demand for radiologists has kept climbing. The Mayo Clinic’s radiology staff has grown by 55% since 2016, to about 400. Hinton himself told The New York Times last year that he had spoken too broadly and was wrong on timing, and that AI would make radiologists “a whole lot more efficient”.
Imaging demand was rising for reasons that have little to do with AI, including ageing populations. But the Jevons logic explains why AI hasn’t reversed that trend: making each read cheaper and faster does not shrink the work worth doing. It makes more of it viable.
Consider a contact centre. Quality assurance teams have typically audited about 5% of calls. That figure was never an optimal benchmark for operational insight. It was a ceiling set by the cost of human reviewers, because checking every call by hand was unaffordable.
Speech AI has turned those economics on their head. Transcribing, indexing and auditing every call now costs a fraction of manual review, so contact centres are moving from spot checks to analysing every interaction: identifying systemic billing problems, tracking customer sentiment and, in insurance, checking that spoken disclosures match what was captured in the CRM system.
In true Jevons fashion, the demand to understand the other 95% of calls was always there. AI simply turned an unaffordable luxury into baseline infrastructure.
Stellenbosch, Soweto, Springs and Saldanha are not Seattle. For South African organisations, the economic shift is sharpest where off-the-shelf global AI falls short. Speech models built in the Global North assume monolingual, high-resource environments. They struggle when conversations switch fluidly between English, isiZulu, isiXhosa and Afrikaans against noisy, real-world backdrops.
Lowering the cost of voice intelligence here means training speech models on small, targeted datasets. Global models are trained on hundreds of thousands of hours of audio; efficient local models need to work with as little as 100 hours of domain-specific recordings. My company, Saigen, builds models to handle code-switching because that is how South Africans speak. Technology should adapt to people, not the other way round.
The better question
South African executives have spent the past few years asking which workforce capabilities AI will replace. The more useful question is: where has demand for operational intelligence been suppressed by cost?
The paradox has limits. It holds where demand is elastic – where a cheaper input unlocks uses that weren’t viable before. Where demand is fixed, cheaper intelligence can simply mean fewer people doing the same work, a real risk for the entry-level roles on which so many young South Africans depend.
The task for executives is to know which kind of demand they are dealing with, and then to work out what becomes possible when the cost barrier falls. Because it will. AI will not extinguish work. It will keep exposing value that businesses could not previously afford to extract.
