The biggest productivity gain is not typing less, but reviewing better.
Fast output changes the bottleneck
AI-assisted development can speed up scaffolding, routine edits and exploration of unfamiliar code. But faster output does not automatically mean a faster release. The bottleneck often moves to testing, integration and confirming that the change fits the rest of the product.
If I cannot explain what the generated code does, I am not comfortable treating it as finished.
Make the task verifiable
Clear specifications, file boundaries, acceptance checks and tests improve the result. For a website change, I want to know whether mobile navigation still works and whether other routes broke. For backend work, authorization and data integrity deserve more than a quick visual review.
Small, reviewable changes are easier to trust than a sweeping rewrite that happens to look impressive.
My take
I see AI as a useful collaborator, not a sign-off authority. I want it to free up time for architecture and problem-solving, but the responsibility for what goes live is still mine.