Been playing around with the idea of using AI to generate workflows from plain English descriptions. The pitch is convincing—describe what you want in natural language, the system generates a ready-to-run workflow, boom, you’ve saved weeks of design and build time.
Tried it on a few real use cases. Some worked surprisingly well. Others generated flows that were technically sound but missed important nuances about how the process actually runs. Things like data quality issues that the plain description didn’t surface, or edge cases that matter in practice but not in theory.
For ROI estimation, this is where it gets tricky. If you feed the generated workflow into your calculator, you’re assuming it’s going to work as designed. But how much of it actually needs rework once real data flows through? The time-saving upside of using AI generation looks great until you factor in the debugging and modification time.
I’m trying to get a realistic picture of where the actual time goes when you’re starting from a plain text description versus a traditional requirements-based approach. Is the savings mostly in the initial design phase, or does a lot of it get reclaimed during testing?
How much of a plain text generated workflow usually survives contact with production data?