How to Keep a Human in the Loop on AI Output Before It Reaches Customers

AI tools write quickly and fluently, and sometimes they get things wrong with total confidence. A draft that reads well can still contain an invented price, a wrong date or a promise the business never meant to make. The fix is a human check before anything reaches a customer. This pattern is called human in the loop, and the glossary has a short entry on human-in-the-loop. This guide shows how to build that check so it works on a busy Tuesday. You end with a risk-based review plan, a checklist and a way to learn from the mistakes you catch.

Step 1: Sort your AI outputs by risk

Not everything needs the same scrutiny. List where AI output is used in your business and rate each use.

Risk level	Examples	Review approach
Low	Internal meeting summaries, brainstorm lists, a first draft you will rewrite	Spot check by the user
Medium	Customer emails, social posts, job ads, proposals	Named reviewer reads every item before it goes
High	Quotes and prices, legal or contract wording, tax, health or safety advice, anything with a binding promise	Subject expert reviews every item, with a checklist

When in doubt, rate it higher. You can relax the rules later when you have evidence. You cannot undo a sent message.

Step 2: Decide who reviews

Name a reviewer for each medium and high risk use. Choose someone who:

• Knows the subject well enough to spot an error.
• Has authority to change or reject the draft.
• Has time to do it properly.

Name a backup for holidays and busy periods. Reviews that depend on one person get skipped when that person is away.

Avoid asking the author to review their own AI-assisted work as the only check. People tend to skim text they just generated. A second pair of eyes catches more.

Step 3: Build a checklist for the common failures

Generic advice like "check it carefully" gets ignored. A short, specific checklist works.

BEFORE SENDING AI-DRAFTED CUSTOMER CONTENT
[ ] Names, spellings and contact details are correct
[ ] Dates, times and locations are correct
[ ] Prices, quantities and GST treatment match our records
[ ] Every fact can be traced to a source we hold
[ ] No promise, guarantee or deadline we have not approved
[ ] Tone fits our business and this customer
[ ] Nothing confidential or from another customer appears
[ ] Links and attachments are right
[ ] I would be comfortable signing this myself

Adapt it to your industry. A builder adds licence numbers and scope exclusions, an accountant adds tax references and a clinic adds privacy checks. Keep it to ten items or fewer, or people will tick without reading.

Step 4: Make facts checkable

Review is faster when the draft points to its sources. Build this into the prompt:

• Tell the tool to use only the facts you provide.
• Ask it to mark anything it is unsure of.
• Ask it to list the figures and dates it used, so you can check them quickly.

For example: "Use only the information below. If a detail is missing, write [CHECK] instead of guessing." Reviewers can then search for [CHECK] and resolve each one. Our guide on writing better prompts for work covers more of these habits.

Be cautious about references and quotes. AI tools can produce plausible-looking sources that do not exist (a failure known as hallucination, explained in what is AI hallucination). Verify any citation, statistic or case before it appears in a document.

Step 5: Put the gate in the workflow

Review works when it is built in, not left to memory. Ways to do that:

• Drafts folder. AI output is saved as a draft, and only the reviewer can send it.
• Approval step in an automation. If you use Zapier, Make or n8n, add a step that sends the draft to a person for approval before it goes out. Automations should never send AI-written customer messages without that step until you have strong evidence of reliability. See setting up an enquiry triage workflow for where the gate sits.
• Sign-off line. Include a short "reviewed by" note in your internal record, with initials and date.

Keep the gate simple. A review step that takes five clicks will be bypassed.

Step 6: Disclose where it matters

Decide when you tell customers that AI was involved. Many routine uses, such as a spell-check or a drafting aid, need no special notice because a person owns the final message. In other cases, such as a chatbot answering customers directly, honesty about what the customer is talking to is good practice. Rules on this can depend on your industry and what you promise. Australia's AI Ethics Principles include transparency and accountability, a useful reference when you decide. Check with an adviser for regulated work. This is general information, not legal advice.

Whatever you decide, the person who sends it is accountable for what it says. For email in particular, see can AI answer customer emails.

Step 7: Log the mistakes you catch

Every error caught in review is useful data. Keep a simple log:

Date	Task	What was wrong	Caught by	Fix

Review the log monthly. Patterns show up quickly: the same wrong figure, the same awkward phrase, the same missing detail. Fix them at the source by improving the prompt, giving the tool better reference material or removing the task from AI.

Step 8: Know when to relax or tighten

After a few weeks you will have evidence. Use it:

• If a task shows no errors across many items and the risk is low, move to spot checks, for example one in five.
• If errors keep appearing, tighten the review, change the prompt or take the task off AI.
• If review takes as long as writing, stop using AI for that task.

Measure how long the full task takes with review included. Measuring whether an AI change is working shows how. The NIST AI Risk Management Framework is a longer reference on managing AI risk if you want depth.

Step 9: Write it into your policy

Put the review rules in your written policy so everyone follows the same standard: which outputs need review, who reviews, and the checklist. See our guide on writing an AI use policy for a template that includes a review section.

Reviewer quick card

1. Read the whole draft. Do not skim.
2. Search for [CHECK] markers and resolve each one.
3. Check every name, number and date against the source record.
4. Remove any promise we have not approved.
5. Read the tone aloud. Does it sound like us?
6. If unsure, rewrite that part yourself or ask.
7. Log any error you found.

What to do next

If you want review routines, prompts and policies set up around the way your team really works, AI adoption from 21Labs builds them into the rollout. It starts with a free discovery. Send the form below and we reply within 24 hours.