A third option that people treat as a failure and which is often the right answer: sell the software with a human attached, on purpose.
The judgement calls you make are real value. There is no rule that says a software business must have no humans in it. A tier where the tool does the mechanical work and you do a review call is a legitimate product, it prices well, and it lets you ship long before the judgement is automated.
What it also does, which is the part people miss: the review calls tell you exactly which judgements to encode next. After twenty of them you will know which three decisions account for most of the value, and those are the ones worth building. Guessing at that from your own experience is much less reliable than watching where you actually intervene.
The honest limits: it does not scale past your calendar, so it is a stage rather than a destination, and you have to be disciplined about reducing the time per customer rather than letting it stay comfortable.
But as a way across the exact gap you are describing, it is the least risky route: you can ship the narrow version from the answer above, attach yourself to it, charge properly, and remove yourself gradually as the tool gets good enough. Nobody is defrauded at any point, which is the thing actually bothering you.