I have a similar story with my washing machine repair. I went through 2 service technicians not being able to diagnose it. Gpt did it and I told the 3rd what to do and it worked.
Yup! That's one of the biggest things we build while shipping this for a Series B SaaS. It's invisible but one of the hardest technical challenges.
We've got guardrails before any code is generated and we put three layers of validation after the feature is ready to make sure it's using the correct data and it's functioning for the use case that the customer asks for.
Our solution to that age old problem is to scope each new feature into its independent "app", which is sandboxed and walled off from the main source code
Yup 80% of the usage is for new frontend features and that's actually quite valuable for our customers who serve traditional enterprises like manufacturers, hotels, labs.
Only some (very well guardrailed) backend functionality is being used.
We're live with five businesses and 2000 daily users! The technical challenges in setting the mapping, ai model, and dsl that creates reliable outputs is the engineering value we deliver.
My shell script definitely wasn't google scale tho!