Chapter 4 · Lesson 2
Add Review Gates Where They Matter
Design human checks around risk instead of reviewing everything equally.
Your team wants AI to save time, but a full manual review of every generated sentence removes most of the benefit.
Match review to consequence
A private brainstorm and a public financial claim should not use the same control. Consider impact, reversibility, sensitivity, and who is affected. Higher-consequence outputs need stronger evidence, specialist review, or explicit approval.
Review the risky fields
Focus attention on facts, calculations, commitments, personal data, legal statements, and decisions. Style can often receive a lighter check. Structured outputs make it easier to route specific fields to the right reviewer.
Make approval visible
A reliable workflow records whether an output is a draft, reviewed, approved, or rejected. Clear states prevent generated material from being mistaken for final work and show who owns the next action.
Build the mental model
Review should be proportional, assigned, and focused on fields where an error has consequences.
Score consequence, not complexity
Consider external reach, sensitivity, reversibility, and cost of correction. A short legal date can need more review than a long internal brainstorm.
Define reviewer authority
Name who may approve, what evidence they need, and whether they can edit or must reject. “Human in the loop” is not a control unless responsibility is clear.
A workflow drafts both internal ideas and customer policy notices.
Have a person quickly review all AI output before use.
Tier 1 internal ideas: owner scans for relevance; no formal approval. Tier 2 customer copy: content owner checks facts and tone. Tier 3 contractual or personal-data content: subject expert verifies source, date, affected group, and wording; named approver records approved/rejected state before sending.
The review effort rises with consequence, risky fields have explicit checks, and the workflow records who can release the output.
Build a lightweight risk matrix
Use two dimensions: likelihood that AI introduces a meaningful error and consequence if the error reaches use. High on either dimension deserves a stronger gate.
Review the matrix after incidents and near misses. A correction that repeatedly takes ten minutes may justify an earlier automated check or a stronger input requirement.
Assign review tiers to three outputs
Classify one low-, medium-, and high-consequence output. Name reviewer, evidence, approval state, and release condition.
Tier: [...] Risky fields: [...] Reviewer: [...] Evidence needed: [...] Release only when: [...]multiple choice
Which generated output requires the strongest review gate?
Choose the answer that best fits the principle from this lesson.
Place the strongest human review where an error would have the largest consequence.