It is the first question every owner asks, and the one the internet is worst at answering. What does this actually cost? Search for an answer and you get a range that runs from "it's free, just use ChatGPT" to a seven-figure transformation programme with a consulting logo attached. Both are true for somebody. Neither is any use to a business turning over between R15 and R50 million, trying to work out whether to spend real money on this at all.
The honest answer is that the shape of the cost has changed over the last three years — and once you see the new shape, it becomes clear that the thing you should actually be nervous about spending money on is not the technology. This post is about both: what AI genuinely costs at your scale, and the far more expensive mistake hiding underneath the question.
The price of the intelligence itself has collapsed
Start with the raw ingredient — the "thinking" the tools do. It has become astonishingly cheap. Stanford University's AI Index, the most respected annual audit of the field, found that the cost of running a model capable of a solid standard of work fell from about $20 to roughly $0.07 to process a million words-worth of text, in the space of about eighteen months — a reduction of more than 280 times. The same report notes that the underlying hardware has been getting cheaper by around 30% a year, and more energy-efficient by around 40% a year.
The practical upshot is blunt: the kind of task that would have cost meaningful money to run in 2023 now costs cents. The intelligence is no longer the expensive part. In many first projects, it is not even a rounding error.
You rent it — you do not buy it
The other half of the story is how you pay for it. You do not build a model, and you do not buy servers to run one. You use models over the internet and pay by usage — closer to how you pay for electricity than to how you would buy a generator. There is no large upfront capital outlay, and the cost scales with how much you actually use, starting near zero.
For an owner, this matters more than the headline price, because it changes the risk of trying. You can start extremely small, prove that something works on one narrow task, and only spend more as it earns more. The cost of a first step is measured in rands, not in a capital budget you have to defend for the next three years.
So what does a first project actually cost?
If the intelligence is nearly free and there is nothing to buy up front, where does the money go? Being honest about the real cost lines is the whole game here:
- The model usage. Now the smallest line, and for a well-chosen first use case, often trivial.
- The plumbing. Connecting the AI to your data and into your workflow so it is genuinely useful, and having someone reliable keep it running. This is a real cost — but at your scale it is one you rent rather than build, because standing up that operational layer yourself would be indefensible for a single use case.
- The one people quietly leave off the estimate. The time to get your data and your process right for that one task, and to check the outputs until you actually trust them. This is real work. Pretending it is free is how projects go wrong.
Add those up and the all-in cost of a well-scoped first project is modest and, crucially, knowable. That is genuinely good news, and it is true now in a way it simply was not three years ago.
The expensive mistake
Here is the part that should reframe the whole question. The money you should actually worry about is not the technology bill. It is the cost of doing this without a clear problem.
The RAND Corporation, in a 2024 study built on interviews with dozens of experienced practitioners, found that more than 80% of AI projects fail — roughly twice the failure rate of ordinary IT projects that do not involve AI. And the root causes were overwhelmingly not technical. The single most common one was a business failing to pin down, or properly understand, the problem the project was meant to solve in the first place. These projects were not sunk by models that were not clever enough. They were sunk by the ordinary business reasons any project fails.
For a large company, a failed AI project disappears into the portfolio. For you, one six-month "AI initiative" with no defined outcome is a serious fraction of your discretionary spend, gone — and that is the genuinely expensive number, not the seven cents per million words. The pattern that wastes money is always the same: start from the technology ("we should be using AI"), build something impressive, then go looking for a use. The pattern that does not waste money is its exact inverse: start from a specific, costly problem you already have, and spend the smallest amount that solves it.
Budget for a win, not a moonshot
In practical terms, that means a few things:
- Cap the spend on a first attempt. Enough to try one narrow problem properly; not enough to hurt if it does not land.
- Insist on a baseline and a number. What does this task cost you in time and money today? What does it cost after? If nobody can answer both, you cannot tell whether you got value — and that is a reason to stop, not to spend more.
- Expand only when the first thing pays for itself. A second project earns its budget when the first has produced something you can point to. Not before.
The reassuring truth is that the cost of trying AI has never been lower, and it keeps falling. The cost of trying it badly — no clear problem, no baseline, nobody accountable — has stayed exactly the same as it ever was. Most of what a good partner is actually worth is keeping you firmly on the right side of that line.
References
- Stanford University Human-Centered AI (HAI) — 2025 AI Index Report (inference cost falling from roughly $20 to $0.07 per million tokens in about 18 months; hardware costs down ~30% a year and energy efficiency up ~40% a year). https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts
- James Ryseff & Anu Narayanan, RAND Corporation — The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (2024): more than 80% of AI projects fail, roughly twice the rate of non-AI IT projects, with predominantly non-technical root causes. https://www.rand.org/pubs/research_reports/RRA2680-1.html
Written by JP Dippenaar, Sixees Labs. Last reviewed July 2026.