Intent Resolution as an External System in AI-Assisted Workflows

This paper examines a foundational architectural flaw in many AI-assisted systems: the conflation of human prompts with executable intent. In common deployments, natural language input is treated simultaneously as an expression of intent and as an instruction to be executed by a large language model (LLM). As AI systems accumulate memory, state, and long-term objectives, this design introduces ambiguity, scope drift, security vulnerabilities, and irreproducible behavior.

The work argues that intent resolution must be externalized from probabilistic models and governed as a deterministic system responsibility. Rather than allowing LLMs to infer intent implicitly from conversational context, reliable systems should translate human input into structured, bounded task representations using explicit state, memory, and policy inputs. The LLM then operates strictly within the constraints of this resolved intent and is not permitted to redefine scope, authority, or objectives.

The paper introduces intent as a first-class system object, distinct from raw user input and distinct from model output. It outlines architectural requirements for external intent resolution, deterministic translation, bounded execution, and explicit handling of memory and state as controlled inputs rather than emergent model behavior.

This work does not prescribe specific schemas, algorithms, or implementations. Instead, it establishes system-level invariants necessary for building AI-assisted workflows that are reliable, auditable, secure, and scalable across sessions, users, and deployments. The findings are applicable to autonomous agents, AI copilots, workflow automation systems, and any architecture where language models interact with real-world processes or long-lived system state.