Hyper-Customized One-Off Software for Cheap

I feel that this essay skips over the way AI lowers the cost of adopting software.

I think the author gets «dependable» wrong: it's not as much low/high level thing, it's more whether making surreal mistakes 1% of the time is amusing or horrifying. However high level an application doing payroll might be, there should be a very clear causal link from the tax laws to what the application does with taxes, at least in cases where tax law can be interpreted without studying precedents! (Garbage-quality parts of the law will need to be interpreted in a manner guided by the user, no way out of it)