Field NotesNext step
Who should review AI output before it goes out?
AI will get things wrong. The question is whether your business catches it or your customers do. Here’s how to build the review step.
Short answer
One named person should review AI output before it reaches a customer, a stakeholder, or a public channel. That person works from a written output standard, a short description of what a passing draft must include and must never say, and has the authority to send a weak draft back. Shared responsibility for review usually means nobody owns it.
Who should be responsible for reviewing AI output?
One person per workflow, by name. Usually it’s the person accountable for the result: the account manager for client emails, the owner for pricing, the sales lead for proposals. Not “the team.”
The AI is not accountable for its output. You are. When a review step belongs to everyone, it gets skipped under time pressure. A named owner turns “someone should look at this” into a step that happens.
What kinds of mistakes does AI make?
AI fails in patterns, not at random. The common ones are tone drift, confident inaccuracy, lost context, over-long answers, and invented details. Because they repeat, you can build checks around them.
- Tone drift. The facts are right, but the voice is too formal, too casual, or too generic for your business.
- Confident inaccuracy. Something wrong, stated with the same certainty as something right.
- Lost context. Over a long session, AI drops constraints you gave earlier and goes generic again.
- Over-completion. You asked for a short summary and got a long one with the point buried in the middle.
- Invented details. Names, dates, and numbers that look real but aren’t. These are the hardest to spot.
What is an output standard for AI work?
An output standard is a short written description of what a passing draft looks like for one task: the information it must include, the tone, words it never uses, and one or two examples of good past work.
Without a written standard, every review is a feeling. With one, two different reviewers reach the same answer, and you can tell whether the prompt or the draft is the problem.
What should an AI review checklist include?
Five questions cover most customer-facing work: is the required information there, does the tone fit the relationship, are any claims unverified, does anything read as obviously machine-written, and is the length right?
- Does it include everything the brief asked for?
- Does the tone fit where this relationship is: new lead, prospect, or long-time customer?
- Are there claims, numbers, names, or dates that need checking before they go out?
- Does anything read as AI-written in a way the recipient would notice?
- Is the length right for the format?
Any “no” stops the draft. The checklist is what keeps review honest. Without it, review means “I glanced at it and it felt fine.”
When should you fix the prompt instead of editing the draft?
When a draft needs heavy editing, uses language outside your standard, or makes a claim you can’t verify, send it back to the prompt. Repeated heavy edits mean the instructions are wrong, not the draft.
Set a rejection threshold before you start. A simple one is a time limit: if editing a draft takes longer than writing it yourself would, stop and fix the prompt or the context. Otherwise your team gets faster at repairing AI output instead of better at producing good output.
Does every piece of AI output need human review?
Anything that reaches a customer, makes a commitment, states a fact, or affects money needs a person’s approval. Internal drafts and notes can get a lighter check, as long as someone owns them.
| The output | Review |
|---|---|
| Customer emails, proposals, quotes | Person approves before it sends |
| Public posts and ads | Person approves before it publishes |
| Reports with numbers | Agent drafts, person checks the figures |
| Internal notes and summaries | Spot-checked by the workflow owner |
What should you do when an AI mistake reaches a customer?
Own it, correct it quickly, and tell the customer what you changed in your process. Don’t blame the tool. Customers care that the work is right, and a visible fix rebuilds trust faster than an apology alone.
Something like: “We found an error in yesterday’s report and corrected it. We also found where our review missed it and updated our checklist.” Then add that type of mistake to your checklist so it gets caught next time.
More questions
What does “human in the loop” mean for AI?
A named person reviews or approves AI work at a set point before it reaches a customer or affects a decision. The loop only works if that person has a written standard and the authority to reject.
Should the person who wrote the prompt also review the output?
They can, for internal work. For customer-facing work, it helps when the reviewer is the person accountable for that relationship, because they know what the customer expects.
How long should reviewing an AI draft take?
A passing draft checked against a written standard should take a few minutes. If reviews keep running long, the prompt or the business context needs work.
Can AI review its own output?
AI can flag gaps and check a draft against a list, which is useful as a first pass. Final approval for anything customer-facing should stay with a person.
Will a review step slow everything down?
Less than fixing a mistake after a customer finds it. A short checklist and a clear owner keep review fast and consistent.