I say on my own homepage that I ship roughly 10× faster than I could a few years ago. That's true, and stated without qualification it's misleading — because the multiplier applies to some of the work and almost none of it to the rest.
Since the honest version is more useful to you than the marketing version, here it is.
What genuinely got dramatically faster
- Anything I've written before. Auth flows, list-detail screens, form validation, API clients, migrations. Work that is well-understood but tedious. This is a large share of any real product and it went from days to hours.
- Reading unfamiliar code. Being dropped into a 200,000-line codebase and needing orientation by Thursday. This might be the single biggest change in my working life — and it's the reason a one-hour audit can now say something useful.
- Tests. The thing everyone under-invests in because it's boring is now cheap enough that there's no excuse.
- First drafts of everything. Not final anything. But the distance from blank page to something to react to collapsed, and reacting is much faster than creating.
- Working across stacks. I'm an iOS engineer by trade. I'm now genuinely useful in backend and infrastructure code in a way I wasn't, because the syntax barrier mostly disappeared.
What didn't change at all
- Knowing what to build. No model has met your customers. The hardest part of product work is deciding what not to do, and that is unchanged.
- Architecture decisions with a two-year horizon. A model will confidently propose a structure that works today and hurts in eighteen months. Telling those apart requires having been burned, which is not a thing you can prompt for.
- Debugging the genuinely hard bug. Race conditions, memory issues, that one crash only on one device model. These still take exactly as long as they always did.
- Anything requiring taste. Whether an interaction feels right. Whether an error message reads as helpful or accusatory. Whether the product is good.
- Coordination. If three people need to agree, AI does not make them agree faster.
What quietly got worse
This is the section missing from most pitches, and it matters if you're buying.
Plausible wrongness scales beautifully. Code that looks right, reads well, and is subtly incorrect is now produced faster than it can be reviewed. The bottleneck moved from writing to verification, and teams that didn't move their attention with it are shipping more bugs, not fewer.
Codebases drift. Generated code tends toward the average of everything ever written, not toward the conventions of your project. Without someone actively holding the line, consistency erodes — and consistency is what makes a codebase cheap to change later.
Juniors are learning differently. I don't know yet whether that's bad. I do know that the understanding I got from being stuck for two days is not obviously replaced by getting an answer in twenty seconds.
What this means for what you're buying
If a supplier tells you AI makes them 10× faster and therefore your project takes a tenth of the time, they're either overselling or they've only ever done the tedious half.
The honest version: the construction phase compressed enormously; the deciding phase didn't move. So the shape of a project changed. It used to be a short thinking phase followed by months of building. It's now a proportionally much larger thinking phase followed by a short build.
Which is why two-week sprints work now and didn't in 2019. Not because engineers got faster at typing, but because the ratio changed — and if you spend the first days deciding properly, the building genuinely does fit in the window.
The question to ask a supplier
Not "do you use AI?" Everyone will say yes.
Ask: "how do you verify what it produces?" The answer tells you whether you're buying compressed delivery or compressed diligence. One of those is worth paying for.
And if you're inheriting a codebase built this way, the audit checklist matters more than it used to, not less.