4 min read
Every day, thousands of businesses rent a Ferrari for the price of a Fiat. They just have not seen the real invoice yet.
The current economics of generative AI are, to put it plainly, artificial. Token prices, the unit cost of every question asked and every answer generated by a large language model, are heavily subsidised by an unprecedented wave of venture capital and hyperscaler investment. Model providers are competing for market share, not margin. The result is that enterprises today consume frontier-grade intelligence at prices that bear little relationship to the true cost of the compute, energy and infrastructure behind it.
This matters enormously for regulated businesses. Health insurers, banks and pension providers operate on structurally thin margins, often constrained by solvency requirements, claims ratios and pricing regulation. Unlike a software start-up, a private health fund cannot simply pass a sudden threefold increase in operating costs through to its members. When the subsidy era ends, and every previous platform cycle suggests it will, the businesses that built core processes on artificially cheap tokens will face an uncomfortable reckoning.
The gap between GenAI enthusiasm and GenAI economics is now well documented. McKinsey calls it the "gen AI paradox": nearly eight in ten companies report using generative AI, yet roughly the same proportion report no significant bottom-line impact (McKinsey, Seizing the Agentic AI Advantage, 2025). Adoption has been broad but shallow, dominated by horizontal copilots whose benefits are spread so thinly across the workforce that they never surface in the profit and loss statement.
Gartner reached a similar conclusion from the cost side, predicting that at least 30 per cent of GenAI projects would be abandoned after proof of concept due to poor data quality, inadequate risk controls, escalating costs or unclear business value. As Gartner analyst Rita Sallam observed, GenAI costs "aren't as predictable as other technologies". A per-token price that looks negligible in a demo becomes a serious line item when multiplied across thousands of customers and millions of interactions in production.
Goldman Sachs Research put the sharpest point on this in its June 2026 report, An AI Job Apocalypse. In a section pointedly titled "AI Agents vs. Humans: Who Costs More?", the analysis shows that the economic case for AI depends entirely on the process it is applied to. MIT's Neil Thompson notes in the same report that capability is not the same as commercial impact, and that reliability, data access and cost remain hard constraints on deployment. The accompanying exhibit comparing agent and human cost curves is worth studying closely, and we would encourage readers to look up the infographic in the original report on the Goldman Sachs website.
The implication is uncomfortable but clarifying: applied to the wrong process, GenAI is more expensive than the humans it was meant to replace, and it does nothing for the top line. A frontier model reasoning through a task that a rules engine, a form or a well-designed decision tree could handle is pure waste, waste that is currently masked by subsidised token prices.
Here is the analogy. Renting a Ferrari at the price of a Fiat feels like a bargain. And yes, you can do the weekly shopping in a Ferrari. It will be fast, impressive and enjoyable. But the shopping does not arrive any fresher, and you could have taken the bus. The moment the rental company starts charging Ferrari prices, the weekly shop becomes ruinous. The intelligent question was never "how much Ferrari can I get?" It was "what does this journey actually require?"
For a health insurer, verifying a policy number, confirming cover levels or triaging a routine claim is the weekly shop. These are deterministic, auditable, high-volume journeys. They demand reliability and compliance, not creativity. Reserving generative reasoning for the moments that genuinely need it, such as understanding a distressed member's free-text description of a complex claim, is where the Ferrari earns its price.
The optimist's rebuttal deserves a fair hearing. Token prices per unit of capability have fallen dramatically year on year, and some argue they will keep falling faster than subsidies retreat, making the "wrong process" problem self-correcting. If intelligence becomes as cheap as electricity, why ration it?
There is truth here, but two caveats apply to regulated businesses. First, consumption grows faster than prices fall. Agentic workflows chain dozens of model calls where a chatbot once made one, so the total cost of ownership can rise even as unit prices drop. Second, cost is not the only constraint. In insurance, every customer interaction carries regulatory weight. A hallucinated benefit explanation is not a cheap error at any token price. Process fit, auditability and predictability remain decisive even in a world of inexpensive tokens.
The lesson for executives in regulated industries is not to avoid generative AI. It is to refuse to let subsidised pricing make architectural decisions on their behalf. Map every customer process, ask what each step genuinely requires, and deploy deterministic automation where determinism wins and generative intelligence where it creates real value. That is how conversational AI increases the top line, through better conversion, retention and member experience, rather than quietly inflating the cost base.
The businesses that thrive after the subsidy era will not be the ones that rented the biggest Ferrari. They will be the ones who knew exactly which journeys needed one, and happily took the bus for the rest.