Human-in-the-Loop AI: 12 Everyday Examples

Human in the loop means the AI prepares and a person decides, at the point where a mistake would be expensive, public or hard to undo. The twelve examples below come from ordinary work: code, support, finance, content, data, hiring, and, with care, medicine and law. Each says what the AI does, where the person steps in, and what happens if they do not. The principles behind choosing those points are in human in the loop for AI agents; this is the list of cases.

1. Reviewing a pull request an agent wrote

• The AI: picks up a card, changes the code, runs the tests and opens a pull request with a summary.
• The person: reads the diff and the evidence before merging. Watch especially for edited tests, new dependencies and changes outside the area the card named.
• If nobody does: a change that passes its own tests but solves the wrong problem, or weakens a test to pass, reaches the main branch and everything built on it.

2. Running a migration on live data

• The AI: writes the migration and its rollback, and runs both on a development copy.
• The person: reads the script and runs it on live, at a time they choose, after a backup.
• If nobody does: an agent with live access and a confident plan can alter or drop data that no later fix restores.

3. Answering a customer

• The AI: reads the ticket and the account history and drafts a reply, citing the help article or the fix it relies on.
• The person: edits and sends. Many teams let routine answers go straight out and route refunds, complaints and anything legal to a person.
• If nobody does: a polite, fluent reply promises something the policy does not allow, or states a cause that is not true, under the company’s name.

4. Approving a supplier invoice

• The AI: reads the invoice, matches it to the purchase order and the delivery, and flags differences in price, quantity or bank details.
• The person: approves payment, and checks any change of bank details with the supplier through a number already on file.
• If nobody does: an invoice with altered bank details, a classic fraud, is paid automatically, because it matched everything except the one thing that mattered.

5. Issuing a refund or a goodwill credit

• The AI: checks the order against the refund policy and proposes an amount with its reasoning.
• The person: approves anything above a set amount, or outside the written policy.
• If nobody does: the policy is applied literally in cases it was never written for, or generously to anyone who asks the right way.

6. Publishing an article or a campaign

• The AI: drafts the piece, suggests headlines and checks links.
• The person: checks every factual claim against a source, reads it for tone, and presses publish.
• If nobody does: an invented statistic or a wrong product claim goes out to every reader at once, and corrections never reach them all.

7. Calling work finished on a task board

• The AI: finishes a task and comments what changed, the commit and how it was checked.
• The person: reads the evidence and moves the card to Completed.
• If nobody does: the finished lane fills with claims rather than results, and the board stops being a record anyone can rely on.

8. Deleting data

• The AI: finds duplicates, stale records or accounts that asked to be removed, and lists them with the reason for each.
• The person: confirms the list and runs the deletion, or lets the AI do it only where the deletion can be undone.
• If nobody does: a pattern that matched ninety-nine right rows also matched one wrong one, and the wrong one was a customer.

9. Granting access

• The AI: reads the request, finds the matching role and prepares the change.
• The person: approves who gets access to what, particularly admin rights and anything touching customer data.
• If nobody does: access creeps upward one reasonable-looking request at a time, including requests an attacker wrote to look reasonable.

10. Screening job applications

• The AI: summarises each application against the published criteria and highlights evidence for each one.
• The person: makes every decision to reject or progress, and looks at the summaries critically rather than ranking by them.
• If nobody does: good candidates are filtered out by a pattern nobody examined. The EU AI Act lists AI used to filter applications↗ (opens in a new tab) and evaluate candidates as high-risk, which brings human oversight duties, and under the GDPR people have the right not to be subject to a decision based solely on automated processing that significantly affects them.

11. Drafting a clinical note

• The AI: turns a recorded consultation into a draft note.
• The person: the clinician reads, corrects and signs it. The record is theirs, not the software’s.
• If nobody does: a misheard dose, a symptom the patient denied recorded as present, or a detail from a different part of the conversation enters a record that other clinicians will trust. This is an example of the pattern, not advice on any product or setting.

12. Reviewing a contract

• The AI: compares a contract with the organisation’s standard terms and marks the clauses that differ.
• The person: a lawyer, or someone with authority to sign, decides what to accept, and checks any authority the AI cites.
• If nobody does: a missed clause is signed, or an argument rests on a source that does not say what the summary claimed. Again, a general example, not legal advice.

What the twelve have in common

Read down the “if nobody does” lines and three kinds of damage repeat: something that cannot be undone (a payment, a deletion, a live migration), something that reaches people outside the team (a reply, an article, a rejection), and something that other work will be built on (a merged change, a record, a finished card). Those are where the person belongs. Everything before them, the reading, matching, drafting and flagging, is where the AI saves the time.

Two habits make the loop work rather than just exist. First, the AI shows its evidence in a form a person can check in a minute: the matched purchase order, the source for a claim, the test that passed. Second, the person’s decision is recorded with their name, so “approved” means someone looked. An approval step that is always clicked without reading is not oversight; it is a delay.

The regulations mentioned above point the same way. Article 14 of the EU AI Act↗ (opens in a new tab) asks that people overseeing high-risk systems can understand the output, decide not to use it, override it and stop the system. That is a useful test for any of the twelve, regulated or not.

Examples 1, 2 and 7 on a board

Several of these examples meet on a task board when the AI is an assistant working your backlog. On fenbs, the four lanes are fixed (To Do, Next Up, In Progress, Completed), and the loop is built from who may move what. The assistant files what it finds into To Do and takes work only from Next Up. When it finishes, it comments its evidence, sets the testing status (Tested, Partly tested, Failed, or Needs owner check when the last check is a person’s), and leaves the card for a person to move to Completed. A live migration gets a card of its own that a person works.

Moving cards is its own permission, so an assistant can be connected with read and comment only and cannot move anything at all. Every change is recorded in History with the assistant’s name and the person it acted for, which is how you check afterwards that the person stepped in where they should have.