What AI should do and what people should keep (Human Review Matrix)

Practicum for Volume 2, Chapter 2, "More Output Does Not Mean More Value." Version: 2026-06. Time: 15-20 minutes. One process on one screen, not a 10-page AI policy.

An outcome does not come from producing more output. It comes from a process with the right review and the right metric. This worksheet turns the chapter's main advice into an exercise: take one process and decide what the machine may do, what needs review, and what a person should keep.

How to fill it in. List 5-7 tasks from one process as rows. Work through the columns for each task. Record the decision in the "Mode" column.

A minimum governance packet is the small set of documents that defines control and ownership. This matrix is the second worksheet. First choose a process with process scoring. Use this matrix to assign a mode to each task. Then define the agent with an Agent Contract and test it on an evaluation set.

Task	Cost of error	Reversible?	Can it be described as a process?	Can you check it?	Mode	Control layer	Flow metric
________	low / medium / high	yes / no	yes / no	yes / no	AI alone / review required / person only
________
________
________
________

What the columns mean:

• Cost of error: what happens if AI gets it wrong and nobody notices.
• Reversible: whether you can undo the action if the wrong result goes out.
• Can it be described as a process? This is the chapter's filter. If the answer is no, the task is not ready for a machine.
• Can you check it? Whether you have a quick way to tell if the result is right, such as a reference answer, rule, or eval.
• Mode: the final decision for the task.
• Control layer for "review required": visibility, guardrails, access limits, or a person at the critical point.
• Flow metric: one end-to-end measure instead of output volume, such as time to result or the share completed without rework.

Default logic

• High cost of error plus poor reversibility means person only, or a person at the critical point.
• If you cannot describe the task as a process, do not hand it over yet. Understand it first.
• If you cannot check the result, build the check first. Otherwise, you have a demo rather than a result.

Output

A completed matrix for one process and one end-to-end metric that you track instead of volume. Return in three months and see what has moved from "person only" to "review required."

---

Next

• For tasks marked "AI alone" or "review required," define the agent with an Agent Contract (Chapter 3).
• Do not send the first version straight to production. Build an evaluation set and run it in shadow mode, where it drafts or observes without taking live action.
• Reviewing output is a scarce skill. Train it with Skills Watch (Volume 1, Chapter 7).
• Follow the full route in the Volume 2 Workbook.

Version: 2026-07-09. What changed: the matrix now has a clear place in the minimum governance packet. The chapter covers flow metrics. This worksheet decides who does what. Warning: check the output outside the model.