Design the Human–AI System

After redesigning the work system, you must specify the boundary between human and AI: what the AI does, what humans do, how they interact, who has authority, and what happens when something goes wrong.

This is not a prompt engineering problem. It is a systems design problem. The AI is one component in a larger human–AI system that must produce outcomes, handle exceptions, maintain accountability, and improve over time.

The question is not "Can AI do this task?" The question is "What should the human–AI system look like, and who is responsible for what?"

For each AI component in the system, specify:

• Input — what information the AI receives, from whom, in what format, with what context.
• Output — what the AI produces, in what form, for whom.
• Authority — whether the AI recommends, drafts, decides, or executes — and under what constraints.
• Confidence and uncertainty— how the AI signals what it knows vs. doesn't know.
• Failure modes — what happens when the AI is wrong, unavailable, or produces harmful output.

For each human role in the system, specify:

• What the human sees — AI output, confidence signals, evidence, alternatives.
• What the human does — review, approve, override, modify, escalate.
• What the human needs — time, competence, context, tools, authority to actually challenge the AI.
• What happens at scale — whether meaningful human oversight is feasible at operating volume.

Every human–AI system needs explicit paths for recovery and escalation:

• Detection — how errors, drift, or unexpected behavior are noticed.
• Override — how humans can intervene when the AI is wrong.
• Escalation — when and how cases move to higher authority.
• Rollback — how to revert to a known-good state when needed.
• Learning — how the system improves from failures.