Workflow, Agent, or Manual Decision Research Note

Workflow, Agent, or Manual Decision Research Note

Question

Product labels do not tell an operator how much autonomy a task should receive. The decision needs to begin with the task, its uncertainty, and the consequence of a wrong action.

Working definitions

A manual process leaves interpretation and action with a person. A deterministic workflow follows declared triggers, rules, branches, and actions. An AI-assisted workflow keeps the flow structure fixed but uses a model for a bounded step such as classification, extraction, or summarization. An agent is given a goal, context, tools, and some discretion over how to proceed.

These are operating definitions for decision-making, not universal industry standards. Many products use "agent" for a flow whose orchestration is largely predetermined.

Google Cloud's agentic design-pattern guide distinguishes predictable sequential work from designs that require model orchestration, custom logic, or human intervention. It recommends human-in-the-loop patterns for subjective or high-stakes tasks that need oversight or final approval.

Workspace Studio illustrates the overlap. Google's conditional-step guidance offers a deterministic Check if step and a model-based Decide step. The surrounding flow can remain fixed even when one decision uses Gemini.

Decision dimensions

The least autonomous model that performs the job well is usually the easiest to test and govern. The choice should consider frequency, input variation, judgment, reversibility, consequence, observability, time pressure, exception rate, and ownership.

High frequency alone does not justify an agent. A repeated task with stable inputs and fixed rules is a strong workflow candidate. Variable language with a bounded, reviewable output may fit an AI-assisted workflow. A goal that requires selecting tools and adapting a plan may justify an agent, but only if the task has enough feedback and control to make that discretion useful.

A manual step remains sensible when volume is low, judgment carries context that cannot be represented safely, the cost of an error is high, or the process is changing too quickly to encode.

Approval placement

Human review is most valuable at a decision boundary, not as a ritual after every low-risk step. Approval should precede money movement, external commitments, deletion, permission changes, publication, regulated decisions, or messages whose recipient cannot be constrained safely.

Review loses value when the person receives too many low-information prompts or cannot see the inputs, proposed action, reason, and consequence. The approval surface is part of the system design.

Editorial use

Use this note to support the decision matrix and four worked examples. The guide should not claim that one category is inherently safe or modern. It should recommend a minimum viable control design and a review date.