In most boardrooms AI is still sold as a magic elixir, promising smarter decisions, happier customers, and fewer staff.
However, as companies move from AI experiments to AI‑everywhere, a token economy is taking shape, and its hidden bill is about to land.
AI token prices have fallen sharply over the past few years with industry analysts putting it down to better hardware and model efficiency. In fact, Gartner predicts that by 2030, performing inference on a large language model (LLM) with 1-trillion parameters will cost GenAI providers more than 90% less than in 2025.
But the business and tech insights company is quick to temper expectations, explaining that frontier intelligence will demand significantly more tokens than current mainstream applications. What’s more, agentic models require between 5-times and 30-times more tokens per task than a standard GenAI chatbot.
“On the surface tokens may be cheaper, but a ‘good’ AI output today usually means bigger context windows, more data stuffed into every request, multiple model calls chained together, and often a higher‑tier model in the loop. So yes, the price per token is down, but the number of tokens per serious task has exploded, and expectations of quality have gone up at the same time,” explains Martin Dippenaar, CEO of Global Kinetic. “The net effect, for real production workloads, is that the true cost per useful answer is often flat, or even higher than it was just eighteen months ago.”
How will this impact local businesses?
For now, Dippenaar says the institutions that are likely to be most exposed to AI’s token economics such as banks, insurers, and telcos are still largely in what he calls the “experimentation and governance” phase.
“Most of the companies that we work with have given their developers free rein with AI technology, which is not the right approach. You need a more measured approach, you need to have governance and guardrails,” he says.
Fortunately, Dippenaar says highly regulated sectors, like banking, lean towards committees, pilot teams and policy documents rather than large‑scale deployments. The result is that the true cost of AI is still mostly theoretical, with models still being tested and not yet embedded into core systems at any significant scale.
Dippenaar believes the compliance, risk, and legacy infrastructure required by banks will insulate them from the worst of the AI bill in the short term. But in the medium term, when they do finally switch on AI at customer scale, the jump in token usage, along with the associated costs, could be brutal.
High-volume customer support companies will feel it first
While banks are still bogged down in pilots and governance, other parts of the economy are already feeling the pain of the current token model, particularly those involved in high-volume customer support, such as call centres.
Today’s call centres handle millions of interactions, all of which are potential candidates for AI agents. Each of these interactions can silently fan out into multiple prompts, long contexts and expensive model tiers, so that a small increase in cost per conversation can turn into a very large monthly bill.
However, Dippenaar says these industries will soon reach a crossover point, where business leaders will need to consider when it becomes cheaper to hire a human rather than have the machines do the work. “As token prices fall but model size and usage intensity rise, providers will be forced to keep the effective cost of an AI agent below that of a human agent. If they don’t, companies in call‑centre‑heavy industries such as airlines, telcos, and banks will find the economics of full automation far less attractive than the marketing pitch of AI,” he says.
That crossover point will arrive sooner than many realise. Gartner has also announced that, by 2028, AI coding costs will overtake the average developer’s salary due to rising large language model (LLM) token consumption and the shift to consumption-based licensing models.
For Dippenaar, that tension lies at the heart of the token economy, where vendors need to keep driving down the unit price of intelligence just to stay ahead of human labour costs, even as customers quietly burn more tokens per task.
The token meter versus owning the engine
Looking ahead, Dippenaar says as AI capability accelerates, intelligence will become a commodity controlled by a handful of firms with the capital to build ever‑larger models and data centres. However, he doesn’t see this as sustainable and he expects a societal, regulatory and business backlash.
With examples like Uber, which Forrester points out, blew through its entire AI budget within the first four months of this year, Dippenaar believes organisations will increasingly look for ways to own more of their AI infrastructure and decouple critical workflows from pure per‑token billing.
“A logical counter‑trend is to bring more of the intelligence engine in‑house by using private or hybrid AI deployments. This will allow the most intensive and sensitive workloads to run on infrastructure the company controls, instead of living permanently on someone else’s meter,” he says. “Companies are right to make the most of AI. We are using it to great effect in our own work, and even guiding clients on how best to optimise it. But it’s vital to remain clear-eyed about the hidden costs, before they wipe out any potential gains.”