Method and System for AI-Based Workflow Generation Using User Context Grounding and Inference-Confidence-Coupled Human-in-the-Loop Placement

Title of Invention

Method and System for AI-Based Workflow Generation Using User Context Grounding and Inference-Confidence-Coupled Human-in-the-Loop Placement

Technical Field

The present disclosure relates to AI-based workflow automation. More specifically, it relates to a method and system that use a large language model (LLM) to turn a user's natural language request into an executable workflow, enrich that workflow by intelligently referring to context information associated with the user, and link the confidence of elements inferred from the context to risk scoring and to the placement of human-in-the-loop (HITL) review points.

Background

With recent advances in LLMs, agent-based workflow automation, which interprets a user's natural language instructions and carries out a series of tasks automatically, is being actively researched. Existing approaches convert a user's natural language request into a workflow graph made up of nodes and edges, then execute it in sequence or with branches.

These approaches, however, have the following limitations.

First, natural language requests are inherently incomplete, and the target of a task, its recipients and its parameters are often left unspecified. Filling these gaps by asking the user again and again or by applying arbitrary default values can reduce both usability and accuracy.

Second, HITL review, in which a person checks and approves the work along the way, is being adopted to make workflow execution more reliable. Reviewing every step lowers productivity, while placing review points by a fixed threshold alone can miss the very inferences that carry the most risk.

Third, because each user's role, history and preferences are not taken into account, the same request can produce a workflow that does not suit the user.