Most teams have now tried an AI coding assistant. Fewer have changed how they deliver software because of one. That gap is worth understanding, because it decides whether the tooling pays for itself.
The bottleneck was never typing
On a healthy team, writing the first draft of a function is a small share of the effort. The time goes on deciding what to build, understanding the system you are changing, reviewing, testing, and releasing without breaking anything. An assistant that makes the first draft faster moves one small part of that.
This is why teams often report feeling faster while their release cadence stays exactly where it was.
Where agents do change the shape of the work
Reading unfamiliar code. Asking an agent to explain a module before you change it is reliably faster than reading it cold.
Mechanical migrations. Renaming across hundreds of files, moving between frameworks, updating call sites after an interface change.
Test coverage for code that already works. Generating cases for existing behaviour is lower risk than generating the behaviour itself.
First-pass review. Catching the obvious before a person spends attention on the subtle.
Where they add work rather than remove it
Generated code is easy to accept and expensive to verify. The failure is rarely a syntax error. It is code that looks correct, passes a shallow test, and is wrong about something the author would have known.
We have seen the same pattern more than once: a feature arrives quickly, then takes longer than usual to stabilise, because nobody on the team can explain why it was written that way.
What to change if you want the benefit
Keep review standards exactly where they were. If anything, raise them.
Insist that whoever submits generated code can explain it line by line.
Measure cycle time and defect rate, not lines produced or suggestions accepted.
Point agents at the work that is tedious and verifiable, not the work that is subtle and consequential.
Used this way the gains are real and unglamorous: less time lost to mechanical work, more attention left for the decisions that actually determine whether the software is any good.