Most firms are now buying AI-built work, knowingly or not
The Office for National Statistics reports that self-reported AI use among UK businesses with ten or more employees rose from around 12% in late 2023 to around 35%. Among the largest firms, those with 250 or more employees, it is 49%. Among the people who build software, use is close to universal: 90% of the technology professionals in Google's 2025 DORA survey use AI at work, as the model's AI app development cost page records.
That raises a plain commercial question that most quotes do not answer. If the work now takes fewer hours, who gets the hours back?
Where the hours fell, and where they did not
We publish a cost model, appdevelopmentcostcalculator.com, that counts the hours for an app two ways: built AI-native, and as a typical agency quote. Each column comes from its own tables, so the gap between them is a result rather than an assumed percentage. Here is its default app:
| Category | AI-native | Agency quote | AI-native agency |
|---|---|---|---|
| Design | 84 | 139 | |
| Build (code) | 121 | 399 | |
| Integrations | 74 | 148 | |
| Code review and hardening | 53 | none | |
| Testing across devices | 109 | 109 | |
| Store release | 26 | 26 | |
| Project time | 64 | 82 | |
| Total hours | 531 | 903 |
Three things stand out. Code is the big saving: 399 hours become 121, because agents write screens, data handling and wiring well. Testing does not move at all, because a phone in a tester's hand takes the same time whoever wrote the code. And a category appears that the old method did not price separately: review and hardening of generated code, 53 hours here. Generated code is fast to produce and still has to be read.
Overall the default app comes to 531 hours against 903. Across the model's eight example apps, from a small utility to a social product, the AI-native column lands between 54% and 61% of the agency column. The saving is real and large, and it is smaller than the claims that code is now free would suggest, because most of an app is not code.
Rates did not fall. Hours did.
If AI made development cheaper, you might expect day rates to drop. The model's review of agency pricing, read on 2 October 2026, found rate cards still in familiar ranges. So the saving, where it exists, is in the hours, and hours are exactly what a fixed-price quote hides.
Who keeps the 372 hours
Take the default app. The difference between the two columns is 372 hours. Where those hours go depends on the billing model, not on the supplier's tools:
Fixed price, quoted on pre-AI hours
You pay for 903 hours. If the work is done AI-native, 372 of them are margin.
Time and materials, billed honestly
You pay for the 531 hours worked. The saving is yours, and so is the risk of overrun.
Fixed price on AI-native hours, unused hours credited
You pay at most 531 hours, and less if the phase finishes early. The supplier carries the overrun risk.
Fixed price is a good contract for a scope you can write down, because the supplier carries the risk of overrun. What goes wrong is a fixed price counted the old way and delivered the new way. The buyer cannot see it, because a fixed price shows a total, not the hours behind it.
Time and materials passes the saving on automatically, if the supplier works AI-native and bills the hours it works. It also passes you the risk. A monthly cap limits that.
The third row is how we price: a fixed price per phase, counted from AI-native hours you can check in the model, with unused hours credited when the phase closes. It is set out in full on how we price.
The bill that grew
One cost went up rather than down: products with AI features now pay a model provider every month. In the model, a chat assistant used by 1,000 people a month costs from $10 to $200 a month in usage depending on the model, and $15.80 to $310.30 with search and content generation added. That bill can be passed through at cost, marked up, or folded into a fee where you cannot see it. Ask which, before you sign. It is one of the five questions we suggest asking any firm.
Questions this essay answers
How much cheaper is software development with AI?
It depends on the category of work. In our published model, code for a standard app falls from 399 hours to 121 when built AI-native, while testing across devices stays at 109 hours and a new review step adds 53. Overall, the AI-native totals come to 54% to 61% of a typical agency quote across eight example apps.
Should a fixed-price quote be lower now that agencies use AI?
Only if the hours behind it were counted the AI-native way. A fixed price estimated with pre-AI hours and delivered with AI tools keeps the difference as margin. Ask how the hours were built up, category by category, and whether unused hours are credited.
Is time and materials better for AI projects?
It passes the saving to you automatically, because you pay for hours worked, but it also passes you the risk of overrun. A monthly cap limits that risk. For a scope that can be written down and signed, a fixed price on AI-native hours with unused hours credited gives you both the saving and the certainty.