For many organisations, AI adoption started small, through proof of concepts and small-scale trials, rather than full-scale rollout.
By Junaid Kleinschmidt, intelligence lead for Accenture, South Africa
Teams tested chatbots, employees began using generative AI tools, and technology leaders explored where the technology could make a difference. AI is now finding its way into more everyday business processes, and as usage grows, so does the bill.
The problem is that the bill is usually much easier to see than the value. A business can measure what it is spending on infrastructure, inference and licences. It is much harder to put a number against the customer that stayed because service improved, the employee who was freed up to focus on more valuable work, or the better forecast that helped release working capital. This creates a gap between the cost of AI and the business value it is producing.
That gap is becoming harder for business leaders to ignore. The source material notes that AI-related spending is tracking to more than $800-billion in 2026, while only 23% of surveyed C-suite leaders report widespread and sustained business value from AI across their organisations. Some companies have also reportedly exhausted a year’s AI budget in a single quarter.
This is why the economics of AI deserve more attention. Tokenomics is essentially about understanding what AI consumes and what the business gets back. Every prompt, response, retrieval step, and agent interaction that leverages an LLM uses tokens, making consumption an increasingly important part of the cost of running AI at scale.
For businesses operating in South Africa, where investment decisions are closely scrutinised and technology budgets need to demonstrate clear value, this is an important shift in thinking. The question should not simply be whether an AI tool is useful. It should be whether the amount being spent on that tool is appropriate for the job it is doing and the value it is creating.
One of the biggest mistakes businesses can make is assuming that every task needs the most capable model available. The source material estimates that only 10% to 20% of enterprise tasks are complex enough to justify frontier or near-frontier capability. Many routine, well-defined tasks can be handled by lighter models, while more sophisticated models are better suited to complex reasoning, long-context analysis or work where errors carry significant consequences.
That distinction matters because AI costs do not behave like traditional software costs. With conventional software, organisations generally know how many licences they have and what those licences cost. With AI, consumption can change according to what people ask systems to do, how much reasoning is required and how much work is being handled by automated agents. As AI becomes embedded in workflows, usage can continue around the clock without a traditional employee or licence count to provide an obvious ceiling.
There is also a behavioural element. Making AI more efficient does not necessarily mean businesses will spend less. As the cost of using AI falls, organisations may simply find more things to do with it. The source material describes this as the Jevons paradox and notes that, in its simulation, only 15% of organisations would bank the savings from a 25% reduction in token prices.
The majority would reinvest those savings in new use cases, larger workloads or higher-quality models. Elsewhere, some organisations are taking a different route entirely: moving to open-source models and smaller, more specialised language models. This does not eliminate cost, but relocates it – instead of paying per token, businesses take on the work of hosting, running and maintaining the infrastructure themselves.
For executives, this changes the conversation. The objective is not to tell employees to use less AI. It is to make sure the organisation is using the right level of AI capability for the work at hand. A routine task that can be completed effectively by a lighter model should not automatically be sent to a more expensive one. At the same time, businesses should be prepared to invest in greater capability where the potential value or risk justifies it.
This requires closer collaboration between technology and business leaders. AI costs often sit within technology budgets, while the benefits appear in customer service, engineering, marketing, finance or other parts of the organisation. Without shared measures, the CFO sees the cost, the CIO sees usage and the business sees the outcome. The source material identifies weak business-technology alignment, insufficient data and analytics capabilities, and difficulty isolating AI’s contribution as significant barriers to linking spending with value.
The practical response is relatively straightforward. Leaders need a common view of AI cost, usage and return. Governance needs to be established before deployments become difficult to manage. Work should be routed to the appropriate level of model capability, and consumption and outcomes should be reviewed regularly as prices, models and use cases change.
For South African businesses, the opportunity is to move beyond treating AI as another technology expense. As adoption grows, understanding where AI is being used, what it costs and what it delivers will become just as important as choosing the technology itself. The organisations that get this right will not necessarily be those that spend the most on AI. They will be the ones that understand where the spend creates genuine business value.