How to decide how much to review any AI output

Executive overview

AI is wrong just enough to cause real damage but right just enough that you stop checking. Reviewing everything kills your leverage; trusting everything leads to costly mistakes. Automation bias means attention degrades as volume rises — so blanket review is not a strategy.

The fix is a single question applied to every AI task: if this is wrong, who pays? That answer determines which of three zones the task belongs to — and how much review it actually deserves.

Route attention to risk, not to habit.

The three zones

• Zone one — cost of error is zero or negligible; no review needed
• Zone two — internal impact; spot-check with binary pass/fail questions
• Zone three — client trust, brand reputation, or legal/financial risk; methodical review required
• Aim for 60–70% of tasks in zone one when starting out
• Most people default to treating everything as zone three, creating a bottleneck

Zone one: skip the review

• Examples: organising files, personal meeting notes, rough research, brainstorming
• Ask: would I proof-read this if I wrote it myself? If no, it's zone one
• Review is not required — the point is to fully delegate and move on

Zone two: binary spot checks

• Examples: emails to your team, internal analysis, training materials, internal presentations
• Spot check = audit, not a methodical read-through
• Write binary questions before reviewing: yes/no, pass/fail only
• Removes subjectivity — a "4 vs 8" judgment call becomes a clear pass or fail
• Example questions: Are the numbers accurate? Do recommendations match the meeting? Is any confidential information included that shouldn't be?
• Speed up by having AI check its own output first: add the binary questions to the prompt with the instruction to revise anything that fails before delivering

Zone three: methodical review with three layers

• Examples: public content, client deliverables, legal or financial documents
• Layer 1 — ground the AI: instruct it to base output solely on provided source documents; if information is missing, say so rather than fill the gap
• Layer 2 — add context: include relevant background (internal politics, audience preferences, stakeholder quirks) that the AI cannot infer
• Layer 3 — ownership check: ask yourself whether you would be proud to put your name on it; if yes, ship it

How zones shift over time

• Trust builds as AI proves consistent on specific tasks — move those tasks from zone three to zone two or one
• Skill development (better prompting, better context-sharing) enables higher-risk tasks to move down
• Model improvements each month expand what AI can reliably do without close oversight
• The goal over time: push as many tasks as possible into zone one to maximise leverage