AI compressed development. The requirements didn't get better — they just landed on QA with no runway left.
Before AI, most of us worked inside the Agile framework - pigs and chickens, skin in the game, stand-ups, retros, story points, the whole vocabulary we all learned to speak fluently, half ironically, half religiously. Development took roughly two uninterrupted weeks per sprint, and inside that window, there was slack. Time to question a product decision, argue about copy, push back on a design choice, go back and forth before anything actually shipped.
That window does not exist anymore.
Part of me wants to resist this. But here is the part I don't like admitting, especially for someone with my trust issues: with AI, we write code faster. We are not necessarily moving faster. Output is up, and code quality has not gone down. The data I track backs that up. That is not the part slipping.
What I am actually seeing slip is the requirements themselves. If copy is not specified, a developer will write their own, and let's be honest, developers are not always the most user-friendly copywriters. The same thing happens with design systems - without an explicit call-out, a developer reaches for whatever ships fastest, and keeping the product coherent is not really their job to catch.
A recent example: we needed one additional checkbox added to our intake flow. Small ask, functionally correct against what was written in the ticket. It still took over two weeks after it landed in QA, because the ticket never addressed what should happen on the negative path, how it interacted with the rest of the flow, what a user actually sees if they check it and then change their mind, and yes, copy - nobody had actually written what the checkbox should say.
This is exactly what I mean by writing code faster without moving faster. The time we saved in development gets spent right back, in the back-and-forth now happening after the fact. We didn't get rid of the slow part. We passed it along, and called that progress.
We used to have time to be wrong. Now we only have time to be right.
That two-week sprint buffer was never really solving the problem in the first place. It was hiding it. We had time to catch sloppy requirements late, so we let them stay sloppy a little longer.
Now that the buffer is gone, the only thing that still works is making the ticket actually ready before it reaches an engineer's desk - copy finalized, components specified, negative paths covered, acceptance criteria someone can actually test against. Not a vague sign-off. A real checklist, enforced before the work starts, not discovered after it ships.
AI didn't just compress the two weeks of development. It compressed the runway we had to catch everything else. Every miss that used to get absorbed quietly inside a sprint is now expensive, visible, and ours to pay for in real time.
