Where AI actually helps in a dev workflow (and where it doesn't)

Every few months a new AI coding tool promises to replace half the job. Most don't. A few genuinely change how fast you can ship. Here's where the line actually is, based on a year of using these tools daily on client work.

Where it earns its keep

• Boilerplate and scaffolding. New CRUD endpoint, a form with validation, a migration file — AI gets these right most of the time and saves real minutes.
• Reading unfamiliar code. Pointing a model at a legacy file and asking "what does this do and why" is faster than archaeology through git blame.
• First-draft tests. Not the whole suite, but the tedious edge-case coverage you'd otherwise skip when you're rushing.

Where it still falls short

• Architecture decisions. Anything that trades off against constraints only you know — budget, team size, what breaks under real load — needs a human making the call.
• Debugging by vibes. AI is confident even when it's wrong. For a gnarly race condition, that confidence is actively unhelpful.
• Product judgment. Knowing which feature not to build is still entirely on you.

The honest framing: it's a very fast junior collaborator, not a replacement for the parts of the job that require actual judgment.

The teams getting the most value aren't the ones chasing every new model release — they're the ones who've figured out exactly which 30% of their workflow to hand off, and stayed disciplined about the rest.