Scott Williams proposes training custom parallel decision models on a company’s historical workflow outcomes, using confidence to handle routine cases and escalating uncertain ones to an existing LLM.
Original by Scott WilliamsAgent workflowsX postOriginal published Source reviewed
Before you dive in
What you’ll find in the original
Historical ticket, invoice, or claim outcomes can form supervision for a bounded model at one recurring workflow step.
Use the cheap model as a confidence-gated preprocessor: confident cases bypass the LLM and uncertain cases follow the current path.
The same pattern could route models or skills, rerank context, screen agent turns, and verify extracted fields.
Worth knowing
This is a prediction and hypothetical architecture, not a deployed result. The suggested “six out of ten” confident cases and spend reduction are illustrative.