Perspective2 min read

Where AI’s work ends and yours begins

A draft can sound convincing and still be wrong. What to review before it becomes a decision.

Prepared with AI assistance · Codex

A magnifying glass over a document with a blue check representing human review
Illustration by Codex for Nadali · CC0-1.0 · 1440 × 960

A convincing draft is not automatically ready to use. Separating preparation from approval helps a team recognize two distinct steps with different responsibilities.

Check claims with consequences

For a customer reply, review timing, price, availability, and promises. For an internal summary, check whether a decision actually happened, who made it, and who owns the next task.

Compare these claims with the original document. Confidence in the wording is not evidence. When a fact remains uncertain, ask the person who can clarify it.

Assign responsibility

Agree on who can use the result. In a small team, this may be the person writing the email or the owner of the relevant process. The tool’s name does not replace accountability for a message that gets sent.

Define situations that need human attention in advance: a customer dispute, an unusual condition, or a request without an approved answer.

Keep a way to correct mistakes

Retain the initial request and final text within your team’s information policies. When an error appears, correct it and revisit the task template. A convenient example should not keep circulating once it has proved inaccurate.

NIST treats risk management as a process across a system’s life cycle. Our practical inference is to make review part of everyday work instead of adding it after a mistake. 

Sources and verification

  1. NIST — Artificial Intelligence Risk Management Framework 1.0 ↗
    Source checked: 2026-10-10