what it's actually good for
frontier models handle the grunt work. the boilerplate, the first draft, the research pass, the repetitive plumbing that used to eat the first three days of a build. that's the honest, unglamorous version of "ai in software development", and it's also the version that actually pays off. a model can write you a hundred lines of correct, boring code in ten seconds. it can read a codebase faster than any human joining a project. it can generate five headline options before your coffee's cold.
none of that is intelligence in the way the marketing implies. it's speed on the parts of the job that never needed a person's judgment in the first place.
what it's still bad at
judgment. taste. knowing when something is technically correct but wrong for the person who's going to use it. deciding what not to build, which is usually the harder call than deciding what to build. an ai will confidently ship you a feature nobody asked for if you don't stop it. it doesn't get bored, doesn't push back, and doesn't know your client the way you do after one phone call.
that gap is exactly where a person earns their fee. the tools got faster. the decisions didn't get easier.
the studio's rule
we sell you the finished thing, not the ai. nobody commissions a build because the AI behind it is impressive. they commission it because the site loads fast, the receptionist answers the phone, the invoice gets sent on time. the model is plumbing. the outcome is the product, and that's the only part a client should ever have to think about.