· 7 min read
The Great AI Reckoning: The Model Is the Cheap Part
AI business cases obsess over model prices, which are now a rounding error. What decides whether AI pays is the people checking its work, and what they miss.
AI cost parity calculator
Will your AI use case pay, once people check its work? Cost a task done by AI against the same task done by a person, including the people who check and fix its output, and find the line it has to stay under.
- AI, per task
- £2.86
- A person, per task
- £2.80
At £2.86 a task against £2.80, AI costs more. It would break even at around 10,796 tasks a month. The model is 0.3% of the cost of each AI task; people checking it, redoing it and living with its mistakes are 72%.
Done by a person£2.80
Done by AI£2.86
- The model’s share of the AI cost
- 0.3%
- People’s share of the AI cost
- 72%
- Most checking time AI can afford
- 1.9 minutes
- Most results it can get wrong
- 7.6%
- Tasks a month needed to break even
- 10,796
- AI cost as a share of a person’s
- 102%
- Saving each month
- −£589
- Saving each year
- −£7,072
Private by designEverything is calculated in your browser. Nothing you enter is sent to us or anyone else.
01The method
How it works
Most AI business cases are built on the price of the model. It is the wrong number. Model prices have fallen faster than almost any cost in history, and they are now a rounding error. What decides whether AI pays is the people around it: the time spent checking its work, the errors they catch and redo, and the errors they miss.
The model
- A person’s cost per task is the time it takes × their hourly cost.
- AI’s cost per task is the model’s token cost, plus the time a person spends checking the result, plus redoing the wrong results they catch, plus the cost of the wrong results they miss, plus a share of the monthly platform costs.
- The two limits show how much room there is: the most time checking can take, and the most results AI can get wrong, before it costs more than a person.
Checking is not a perfect safety net. People tend to trust AI output, and the more plausible it looks, the more errors get through. A large review of human–AI experiments found that, on decision tasks, people working with AI often did worse than the better of the two working alone. If checking a result takes nearly as long as producing it, AI cannot pay at any volume.
Before it gets the green light
- Which measure will it improve, by how much, and by when?
- How long does checking really take, measured rather than estimated?
- What does a mistake that gets through cost, and who carries it?
- Could a smaller model, a simple rule or a better process do the job for less?
Sources
- Stanford HAI (2025). AI Index Report 2025. The cost of GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024.
- Dratsch, T. and others (2023). Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology. When the AI was wrong, very experienced readers’ accuracy fell from 82% to 45.5%.
- 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. 106 experiments; human–AI combinations did worse than the best of either alone on average, with losses concentrated in decision tasks.
- METR (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity. Developers took 19% longer with AI, while believing it had made them faster.
- Office for National Statistics (2025). Annual Survey of Hours and Earnings. Median full-time hourly pay of £19.67.
From the article
Want a second opinion on your numbers?
Tell us what the tool showed you. We will help you test the assumptions and work out a practical next step.
02Tools