The problem
Every company has a routine multi step process that eats hours and follows rules somebody could write down. It is the most obvious candidate for automation there is.
It is also where naive LLM automation does real damage, because a plausible wrong answer executed automatically is far worse than no automation at all.
What I built
- Prompt templates per workflow step, so each step has one narrow job instead of one prompt trying to do everything at once.
- Output validation between steps, so a malformed or implausible result stops the chain rather than feeding the next step bad input.
- Human in the loop checkpoints at the steps where being wrong is expensive, and full automation where it isn't.
The hard part
Deciding which steps a human still has to see. Automate too little and nobody saves any time. Automate too much and the first bad output goes straight through to a customer with nobody watching.
Outcome
In progress
Building on the agent already shipped at Panelist
Validated between steps
Bad output stops the chain
Human where it counts
Checkpoints on the expensive steps
Stack
In build
Workflow Automation Agent
Turn a multi-step manual process into a guarded AI workflow with validation and human-in-the-loop checkpoints.
Have a similar problem?
Thirty minutes, no pitch. If this isn't the right approach for you, I'll say so.