There is a version of the AI story where a solo founder ships a company with no team. There is a more sceptical version where none of it works. Neither matches what we see when we actually build with these tools every week.

What it genuinely does

The compression is real. A first working prototype that used to take a small team weeks now takes a capable person days. Boilerplate, scaffolding, test data, first-draft copy, the tedious integration glue between two APIs: all of it is dramatically faster. For validating whether an idea is worth pursuing, that is a serious advantage.

It is also a genuine substitute for the specialist you cannot yet afford at 11pm. A founder who is not a designer can get to a defensible interface. A founder who is not an engineer can get far enough to know whether the hard part is hard.

What it does not do

It does not decide what to build. Judgement about which customer problem is worth solving comes from talking to customers, and no model has your market.

It does not carry accountability. When a system mishandles real money or real medical data, someone has to own that, and it cannot be a model. Code that looks correct and is subtly wrong is the failure mode we are called about most, and reviewing it well requires the expertise the tool was supposed to replace.

It does not maintain anything. Generated code becomes your code the moment it ships, and the second year of a system costs more than the first regardless of who typed it.

The useful framing

Treat AI as an exceptionally fast, tireless junior who has read everything and understood the stakes of none of it. Enormously valuable with direction and review. A liability without either.

The founders getting the most from it are not the ones asking it to replace a team. They are the ones using it to get to a real answer faster, then bringing in people who know what they are looking at before it touches production.