AI is very good at some parts of building software and actively unhelpful at others. Knowing which is which is most of the value.
Where it genuinely helps
Getting from blank page to draft. The hardest part of a spec, a schema or a landing page is the first version. A draft you disagree with is far more useful than an empty file, because disagreement is specific.
Exploring options fast. "Show me three ways to model this" costs a minute instead of an afternoon.
The tedious middle. Converting a payload shape, writing the fourth similar endpoint, generating test fixtures — mechanical work with a clear right answer.
Explaining unfamiliar code. Pasting a function you inherited and asking what breaks if you change it is quicker than reading upward through five call sites.
Where it costs you time
Anything load-bearing you cannot check. If you cannot tell whether the output is right, you have not saved work — you have deferred it to an incident.
Decisions that need context it does not have. Why your team picked this queue, what the client promised, which customer is about to churn. Those are not in the prompt.
Code you will not read. Accepting a large block you have not understood is borrowing against future debugging time, at a bad rate.
Use it to go faster on things you can verify. Use your own judgement on things you cannot.
A workflow that holds up
- Describe the problem, not the solution. You get better answers by explaining the constraint than by asking for a specific implementation.
- Ask for the shape first. Types, function signatures, the schema. Review those before any body is written — a wrong shape reviewed early costs nothing.
- Make it justify itself. "What breaks if this runs twice?" surfaces more than "is this correct?".
- Keep the human in the loop where it matters. Anything touching money, auth or data deletion gets read line by line.
The compounding part
The real acceleration is not typing speed. It is that the cost of exploring an idea drops far enough that you explore more of them. Teams that used to evaluate one approach now sketch three and pick better.
That is worth more than any amount of generated boilerplate.