Course outline for Human–AI Workflow Design

Module 1: Allocate work and authority

• Human-agent collaboration: Assign drafting, recommending, approving, and executing to named roles with explicit evidence requirements and decision responsibilities.
• Division of work: Allocate tasks according to consequence, reversibility, and judgment needs rather than assigning everything a model can generate.
• Human approval: Choose pre-action review for consequential decisions and require evidence checks rather than treating an approval button as sufficient.
• Human escalation: Define who receives missing, contradictory, or unsupported recommendations and what evidence enables a responsible next decision.

Practical outcome: Create a draft/recommend/approve/execute matrix and escalation rules based on consequences and reversibility.

Module 2: Make review an active control

• Authority boundaries: Separate recommendation privileges from execution rights and deny simulated sends that lack the required human authorization.
• Uncertainty: Identify missing support and contradictory evidence without substituting persuasive wording or stated model confidence for justification.
• Review: Apply independent evidence checks and acceptance criteria while measuring reviewer time, disagreements, and available review capacity.
• Skill retention: Design practice tasks that preserve the domain judgment reviewers need to recognize errors and challenge automated recommendations.

Practical outcome: Define independent evidence checks, reviewer-capacity limits, and practice tasks that preserve judgment.

Module 3: Test and revise collaboration

• Simulator-based practice: Use supplied correct and persuasive incorrect outputs to rehearse accept, reject, and escalate decisions without real consequences.
• Automation bias: Detect unjustified acceptance of convincing recommendations and compare reviewer decisions with independently checked supporting evidence.
• Ironies of automation: Examine how reduced routine practice and concentrated exception handling can weaken oversight, then revise one bounded collaboration control.

Practical outcome: Run persuasive correct and incorrect cases, record reviewer disagreements, and propose a bounded policy correction.

Practical exercise

• Produce an authority matrix, review checklist, and escalation record for a fictional draft-and-approve process.
• Review six supplied outputs, including correct, unsupported, and contradictory recommendations; justify accept, reject, or escalate decisions with evidence.
• Verify that simulated unapproved sends are denied and unsupported recommendations cannot pass the acceptance checklist.
• Measure reviewer time and disagreement, then revise one control and specify a skill-retention exercise.