Clients have started asking whether AI makes projects cheaper. The honest answer is: some line items, substantially; the total, less than you would hope.
What actually got faster
Boilerplate, CRUD layers, data mapping, migrations, test fixtures, first drafts of documentation, and porting a known pattern into a new module. On these, agent-assisted work is dramatically quicker and the quality is fine because the target is well-defined.
What did not
- Deciding what to build. Discovery takes the same time it always did.
- Reviewing. More code arrives, so more code must be read.
- Integration work against third-party services with undocumented behaviour.
- Anything touching money, permissions or personal data, where the bar for verification is the same as it ever was.
How we estimate now
We split the work into "well-specified and pattern-following" and "requires judgement", and apply different assumptions. The first bucket shrank. The second did not. On a typical project the total lands somewhere between 10% and 25% below where it would have been — real, but not a different order of magnitude.
The bottleneck moved from typing to deciding. Estimates should follow the bottleneck.
The trap
Pricing as if everything got faster produces a schedule where review is the unfunded activity. That is how teams end up with a large codebase nobody understands and a defect rate that climbs quietly for two quarters.