The wrong question
Whether AI can write production code is no longer an interesting question. It can, it does, and it is getting better every quarter. Teams that still debate this are answering last year's exam.
The question that decides whether your product ships well is a different one: where. Where does AI make an engineer faster, and where does it quietly make a product worse while looking perfectly confident about it?
Where AI earns its place
First drafts of interfaces. Test suites. Data migrations. Glue between APIs. The thousandth variation of a form, a table, a webhook handler. This is most of the volume of software work, and AI handles it at a pace no team of humans can match.
Used this way, the gain compounds. An engineer who spends the morning reviewing and steering instead of typing gets through a week of conventional work before lunch. That is where our speed comes from. Not heroics, allocation.
Where it doesn't
Data models you will live with for years. Security boundaries. Payment flows. Architecture choices whose cost only shows up at scale. Here the danger is precise: a plausible answer and a correct answer look identical on the screen. Only judgment built from shipping real products can tell them apart, and AI does not carry the consequences of being wrong. We do.
How we work because of this
Every line that ships passes through a senior engineer who understands the whole system, not just the diff. AI accelerates the work. Engineers own the outcome. That split is not a compromise between old and new ways of working. It is, as far as we can tell, simply how good software gets built now.