The discussion surrounding artificial intelligence has reached a state of tedious equilibrium. Company leaders are told either that AI will lead to catastrophic job cuts or that it will finally deliver promised operational efficiency. Sometimes they are told both.
Each side shares an unfounded assumption: that corporate demand for intelligence—that is, the ability to solve problems—is constant. However, history shows otherwise. When the cost of a vital resource falls, demand rarely remains stable; it often increases, and significantly so.
Artificial intelligence lowers the unit cost of information processing, and with this, organizations' appetite for intelligent solutions grows. In 1865, English economist William Stanley Jevons published 'The Coal Question,' in which he made a counterintuitive observation. Improvements to the steam engine, particularly those developed by James Watt, sharply reduced the amount of coal needed to perform a specific task: a fully developed Watt engine consumed about a quarter of the fuel compared to Newcomen designs.
It was intuitively assumed that Britain would burn less coal. But the opposite happened. Cheaper work made steam power applicable in thousands of new areas—from textile mills to deep mines and railways. Coal consumption soared. This phenomenon became known as the Jevons Paradox and today serves as an economic driver for AI. AI does not eliminate corporate demand for intelligence; it reduces its unit costs. When applying intelligence to solve a problem becomes cheaper, organizations do not reduce its use; they activate a huge reserve of latent demand—problems that always existed but were too expensive to solve.
Ten years ago, radiology seemed like the perfect candidate for AI displacement. Image recognition models could analyze thousands of scans tirelessly. In 2016, Geoffrey Hinton, who later received the Nobel Prize, stated: 'I think if you work as a radiologist, you are like a coyote that has already crossed the edge of a cliff but hasn't looked down yet. People should stop training radiologists now. It is obvious that deep learning will surpass radiologists in five years.'
The situation turned out to be different. Radiology became a primary target for AI application in medicine—about three-quarters of the more than a thousand AI applications approved by the U.S. Food and Drug Administration for medical use relate to radiology. Nevertheless, the demand for radiologists continues to grow. The pool of radiologists at Mayo Clinic increased by 55% since 2016, reaching approximately 400 people. Hinton himself told The New York Times last year that he had spoken too broadly and was wrong about the timeline, noting that AI would make radiologists 'much more efficient.'
The rise in visualization needs was driven by factors largely unrelated to AI, including an aging population. However, Jevons' logic explains why AI has not changed this trend: making each analysis cheaper and faster does not reduce the volume of work that needs to be done. It makes more work feasible.
Consider the contact center. Quality control teams usually checked only about 5% of calls. This figure was never an optimal indicator for obtaining actionable intelligence. It was a ceiling set by the cost of human reviewers, as checking every call manually was prohibitively expensive. However, voice AI has radically changed this economic model. Transcribing, indexing, and auditing every call now costs only a fraction of manual review, leading contact centers to move from spot checks to analyzing every interaction: identifying systemic billing issues, tracking customer sentiment, and, in insurance, cross-referencing verbal claims with data entered into the CRM system.
In true Jevonian spirit, the need to understand the remaining 95% of calls has always existed. AI has simply transformed an unaffordable luxury into basic infrastructure.
For organizations in South Africa, such as Stellenbosch, Soweto, Springs, and Saldanha, the economic shift is most palpable where universal global AI fails. Voice models created in the Global North assume a monolingual environment with high resources. They struggle when conversations fluidly switch between English, isiZulu, isiXhosa, and Afrikaans amidst real-world noise. The reduction in the cost of voice intelligence here is achieved by training speech models on small, targeted datasets. While global models are trained on hundreds of thousands of hours of audio, effective local models can operate with just 100 hours of domain-specific recordings. The company Saigen creates models for code-switching because that is how South Africans speak. Technology must adapt to people, not the other way around.
A more important question. Leaders in South Africa have spent the last few years wondering which job functions AI will replace. A more useful question is: where has the demand for operational intelligence been suppressed due to cost? The paradox has its limits. It operates where demand is elastic—that is, when a cheaper resource opens up possibilities that were previously unviable. Where demand is fixed, cheaper intelligence may simply mean fewer people doing the same job, which poses a real risk to entry-level positions upon which many young South Africans depend.
The task for leaders is to determine what type of demand they are dealing with and then figure out what becomes possible when the cost barrier drops. Because it will happen. AI will not destroy jobs; it will continue to uncover value that businesses could not previously afford to extract.