What the human still decides

People ask us, reasonably, what’s actually left for a person to do once a model can write code, draft copy, and answer support tickets. Here’s our list, the one we actually use.

Scope

A model will happily build whatever you ask it to build, including the wrong thing, built well. Deciding what a feature should do, and just as often what it shouldn’t, is a business decision dressed up as a technical one. We don’t outsource that. A client tells us the problem, we decide the shape of the fix, and the model helps us build whichever shape we picked.

Anything with a legal or financial edge

Reconciliation logic, patient data, a payment flow, an access rule that decides who can see what. These are places where a plausible-looking answer that’s subtly wrong causes real damage. A model can draft the first pass of the code. A person reads every line of it before it touches real money or real records, and writes the tests that would catch the mistake the model didn’t think to make.

Taste

This is the hardest one to defend on a spreadsheet and the easiest one to feel when it’s missing. Which of three drafts sounds like a person wrote it. Whether a screen is calm or just empty. Whether a joke in a support message lands or falls flat. A model can generate options all day. Picking the right one, for this client, in this moment, is still a person’s job, and we think it always will be.

What happens when the model is wrong

Every model is wrong sometimes, confidently, in ways that read as correct. The question isn’t whether that happens. It’s whether anyone would have caught it. Our answer is that a person reviews the output the same way they’d review a colleague’s work: reading it, not skimming it, checking the part that seemed too easy.

Why we say this out loud

We could keep this list to ourselves and let clients assume the model does more or less than it actually does. We’d rather say it plainly, project by project, because a studio that won’t name where the model stops isn’t one you can really trust with the parts that matter.

The short version

If a mistake here costs a client money, trust, or a compliance problem, a person made the final call, not a model. Everything else is fair game for the model to draft, iterate on, and speed up. That line moves as tools get better. Right now, this is where we draw it, and we tell clients exactly where, not as a disclaimer but because it’s the honest answer to a fair question.