What AI workflow design actually involves
AI workflow design is the practice of breaking a piece of work into steps, deciding which steps an AI tool handles and which a person handles, and defining how work passes between them. Concretely, a workflow designer writes down the steps of a recurring task, marks where drafts get generated, where a human checks them, what good enough to pass means at each gate, and what happens when the AI output fails the check. The output is a documented, repeatable process, not a clever prompt.
The skill borrows from process design that predates AI. O*NET task descriptions show that many occupations already contain organizing, scheduling, and process-improvement tasks; workflow design extends those with AI-specific decisions. The distinctive new questions are about failure: where can a fluent but wrong output slip through, what is the cost when it does, and which checkpoints catch it cheaply. A designed workflow makes those answers explicit, so quality does not depend on whoever happens to be careful that day.
Why this skill compounds as AI spreads
The mechanism is leverage. A person who uses AI well improves their own output. A person who designs workflows improves the output of everyone who follows the workflow. As AI tools spread through an organization faster than anyone's ability to use them consistently, the gap between ad hoc use and designed use widens, and closing that gap is exactly what this skill does. The value scales with adoption: more AI use in the organization means more processes needing design.
Observational evidence fits this picture. Anthropic's Economic Index work describes usage that leans toward augmentation, with people iterating alongside AI on tasks rather than handing work over wholesale. Augmentation at scale requires someone to decide how the human and the tool divide each task, which is workflow design by another name. The ILO's transformation framing points the same direction: if most jobs change shape rather than disappear, someone must design the new shape. That someone is currently scarce in most teams, though how long the scarcity lasts is unknown.
Building the skill in 90 days
Days 1 to 30: map one workflow you already own. Pick a recurring task that takes you at least an hour, write out every step, and time each one for two weeks. Then mark each step as draft-suitable for AI, judgment-required, or verification. This map is your first artifact, and building it teaches you the core diagnostic skill: seeing work as steps with different properties rather than as a single blob.
Days 31 to 60: redesign and run it. Introduce AI drafting at the steps you marked suitable, define a pass/fail check for each handoff, and run the redesigned workflow for a month while logging time spent and errors caught. Days 61 to 90: write the playbook. Document the workflow so a colleague could run it without you: steps, tools, checks, and known failure modes. Then have one person actually try it and revise based on where they got stuck. The playbook plus your before-and-after log is your portfolio piece.
