Check out this blog post, “Beyond Prompt-and-Pray” by Hugo Bowne-Anderson and Alan Nichol that talks about how the current approach of just letting AI handle everything (prompt and pray) is unreliable and expensive.

The “prompt and pray” approach has serious limitations once you get out of the prototyping phase and need to provide a consistent experience.

Complex workflows require more control than simply trusting an LLM to figure everything out… Debugging these systems is a nightmare…

Agreed. Reviewing transcripts and resorting to trial and error simply doesn’t scale. In my current role, the conversational design is very important and we want to keep the user on track with the program we’ve created for them. Combining AI’s understanding of natural language with clear and defined workflows makes things reliable, secure and safe.

Structured automation is a development approach that separates conversational AI’s natural language understanding from deterministic workflow execution.

Fully autonomous agents don’t provide the necessary oversight and safety that some applications need. Your trusted application with a focused, thoughtful user experience will require predefined, testable workflows that keep the business logic separate from conversational capabilities.

The future of enterprise conversational AI isn’t in giving models more runtime autonomy—it’s in using their capabilities more intelligently to create reliable, maintainable systems.

And this is why I’m not worried about AI taking engineering or product management jobs any time soon.

Check out the blog post for more detail on what a structured automation should look like. Bowne-Anderson and Nichol do a nice job of providing some real work examples to explain their points. Great read.