Almost every AI business case starts with the price of the model: so much per million tokens, so many tokens per task. It is the wrong place to start. Model prices have fallen faster than almost any cost in the history of computing, and for most business tasks they are now a rounding error.
What decides whether AI pays is everything around the model: the people who check its work, the mistakes they catch and fix, the mistakes they miss, and the platform it runs on. Those costs are rarely in the pilot, and they are often the whole story. This article sets out what the evidence says, and there is a calculator to test your own use case.
Adoption is high; returns are not yet
Use of AI has spread fast. Stanford’s AI Index reports that 78% of organisations were using AI in 2024, up from 55% a year earlier, and 88% by its 2026 edition. But the returns organisations report are modest: in the survey data the 2025 report draws on, most reported cost savings were under 10% and the most common revenue gain was under 5% Stanford HAI.
Analysts expect a shake-out. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing “poor data quality, inadequate risk controls, escalating costs or unclear business value” Gartner, and that over 40% of agentic AI projects will be cancelled by the end of 2027 for the same reasons Gartner. These are forecasts, not measurements.
You may have seen the claim that 95% of AI pilots deliver nothing. It comes from a 2025 MIT report based on a small, non-random sample, which defined success as sustained profit-and-loss impact, and it has been widely criticised for its methods HPCwire. The honest position is less dramatic and more useful: AI often works, it often doesn’t pay, and the difference is rarely the model.
The model is the cheap part
The Stanford AI Index found that the cost of using a model as capable as the original GPT-3.5 fell from $20 to $0.07 per million tokens between November 2022 and October 2024, a drop of more than 280 times Stanford HAI.
At today’s prices, a typical business task (reading a few pages, writing a short answer) costs well under a penny in tokens. In the worked example below, the model is 0.3% of what each task done by AI really costs. Halve the model price, or double it, and the answer barely moves. Negotiating hard on token prices while ignoring everything else is optimising the wrong number.
The people are the expensive part
Checking takes time
Most serious uses of AI keep a person in the loop, and they should. But checking is work. If a person takes six minutes to do a task and two minutes to check the AI’s version of it, a third of the saving is gone before anything has gone wrong. If checking takes as long as doing, AI cannot pay at any volume, however cheap the model.
Checkers miss things
A person in the loop is not a perfect safety net, because people tend to trust automated suggestions. In a study of radiologists reading mammograms with AI suggestions, when the AI was wrong, the accuracy of very experienced readers fell from 82% to 45.5%, and inexperienced readers fared worse Radiology. A meta-analysis of 106 experiments found that, on average, people working with AI performed worse than the better of the person or the AI working alone, with the losses concentrated in decision-making tasks and gains in creating content Nature Human Behaviour.
So every AI business case needs two numbers it usually lacks: how often the AI is wrong, and how many of those errors get past the person checking. The errors that get through carry a cost, and someone outside the pilot usually pays it.
Feeling faster is not being faster
In a randomised trial, METR found that experienced open-source developers took 19% longer to complete tasks when allowed to use AI tools. Before starting, they expected AI to speed them up by 24%; afterwards, they still believed it had sped them up by 20% METR. The study was small and used early-2025 tools, and METR has since said developers are likely faster with AI in 2026, while cautioning that its newer data is weak evidence of how much METR.
The lesson is not that AI slows people down. It is that people’s sense of their own productivity is an unreliable measure. The UK’s cross-government trial of Microsoft 365 Copilot found self-reported savings of 26 minutes a day GOV.UK; where departments measured against a comparison group, the figures were more modest. Measure; don’t survey.
Where AI genuinely pays
None of this means AI doesn’t work. It means it works unevenly.
- In a study of over 5,000 customer support agents, access to a generative AI assistant increased issues resolved per hour by about 14–15% on average, and by 34% for novice and less-skilled workers, with little effect on the most experienced NBER QJE.
- In an experiment with 758 BCG consultants, those using AI on tasks within its capabilities completed 12.2% more tasks, 25.1% faster, with more than 40% higher quality. On a task just outside its capabilities, they were 19 percentage points less likely to get the right answer than those working without it HBS working paper.
The researchers call this the “jagged frontier”: AI is very good at some tasks and confidently wrong at others that look similar. The value is in knowing which is which, for your work, before you scale.
Scale matters too. A study by MIT FutureTech and IBM of computer vision tasks found that only 23% of the wages paid for vision tasks would be economically attractive to automate; for the rest, people were cheaper, largely because the systems only pay off at volumes most single organisations never reach MIT.
A worked example
A task takes a person six minutes, at a cost of £28 an hour including employer costs (UK median pay was £19.67 an hour in 2025 ONS). The organisation does 10,000 a month. With AI, a person spends two minutes checking each result; 8% of results are wrong, the checker catches 70% of those, and a mistake that gets through costs £40 to put right. The platform costs £8,000 a month.
| Per task | Cost |
|---|---|
| A person doing it | £2.80 |
| Model | under 1p |
| Checking | £0.93 |
| Redoing errors caught | £0.16 |
| Errors that get through | £0.96 |
| Platform share | £0.80 |
| AI, all in | £2.86 |
AI costs slightly more than doing the task by hand. The model is 0.3% of that; people checking, redoing and living with its mistakes are 72%. Two numbers decide the outcome: checking must take less than 1.9 minutes, and the AI must be wrong less than 7.6% of the time. Neither has anything to do with the price of the model.
Change the assumptions in the calculator, and you will find that small changes in checking time and error rates swing the answer far more than any change in model price.
Before you scale
- State the value. Which measure will it improve, by how much, and by when? No green light without an answer.
- Measure checking time. Time it with real people on real work, not in a demo.
- Measure errors, and the ones that get through. Test with known answers. Then ask what an error costs when it reaches a customer, a decision or a record.
- Find the edge of the frontier. Test the tasks that look like the ones AI does well but aren’t.
- Right-size the tool. A smaller model, a simple rule or a better process may do the job for less, with fewer surprises.
If you would like an independent view of an AI business case before it scales, talk to us. We will tell you if a simpler answer would do.
Sources
- Stanford HAI (2025). AI Index Report 2025, including the economy chapter.
- Gartner (2024). Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025.
- Gartner (2025). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027.
- Dratsch, T. and others (2023). Automation bias in mammography. Radiology, 307(4).
- Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful: a systematic review and meta-analysis. Nature Human Behaviour.
- METR (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity, and the 2026 update.
- Government Digital Service (2025). Microsoft 365 Copilot experiment: cross-government findings report.
- Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2); NBER working paper 31161.
- Dell’Acqua, F. and others (2023). Navigating the jagged technological frontier. Harvard Business School working paper.
- Svanberg, M. and others (2024). Beyond AI exposure: which tasks are cost-effective to automate with computer vision? MIT FutureTech research brief.
- Office for National Statistics (2025). Annual Survey of Hours and Earnings.
- HPCwire (2025). MIT report flags 95% GenAI failure rate, but critics say it oversimplifies.
Want to talk this through?
We help organisations put ideas like these into practice. Tell us what you are working on and we will help you find a practical next step.
