Human in the loop examples: where the person decides
Human in the loop examples from everyday work: what the AI drafts, what the person decides, and where to put the approval step in your own workflow.
On this page 7 sections
Human in the loop means an AI system drafts, suggests or prepares, and a person checks and decides before anything consequential happens. The examples below show where that check sits in ordinary work: what the AI does, what the person does, and what would go wrong without the person. Use them to decide where your own approval step belongs.
For the design principles behind approval steps, read human in the loop: let AI draft the change, keep the decision. This post is the practical companion: examples you can copy. Download Plexii free for Mac or Windows if you want an assistant that asks before it changes your work.
The pattern in one line
Every example below has the same shape:
AI drafts → a person reviews the real output → the person approves, edits or rejects → only then does anything change.
Stanford's Institute for Human-Centered AI describes human-in-the-loop systems as ones that "include human feedback or intervention as part of their operation" (opens in a new tab), where people may guide, correct errors or make the final decision. The EU's AI Act, Article 14 (opens in a new tab), requires high-risk AI systems to be designed so they "can be effectively overseen by natural persons", and warns about automation bias: the tendency to over-rely on what the system produces. That warning applies to everyday tools too. An approval step only helps if the person actually reads what they approve.
Human in the loop examples at work
| Work | What the AI does | What the person decides | Risk without the person |
|---|---|---|---|
| Client email | Drafts a reply from the thread | Reads it, fixes tone and facts, sends it | A wrong promise goes out under your name |
| Meeting actions | Lists actions and owners from notes | Confirms each owner agreed to it | Someone is assigned work they never accepted |
| Project table | Drafts rows: tasks, dates, owners | Checks dates and owners, deletes weak rows | A plan that looks finished but is not real |
| Research summary | Summarises sources with citations | Opens the cited passages before quoting | A quote or figure the source does not contain |
| Spreadsheet formula | Writes a formula from a description | Tests it on a few rows with known answers | Silent errors across thousands of rows |
| Invoice coding | Suggests account codes for line items | Approves the batch, queries the odd ones | Expenses filed against the wrong account |
| Job applications | Summarises CVs against criteria | Reads the CVs for the shortlist decision | Good candidates screened out unseen |
| Customer support | Suggests a reply and a category | Sends, edits or escalates | A confident wrong answer to a customer |
| Calendar planning | Proposes blocks for the day | Moves, drops or accepts the plan | A day planned around the wrong priority |
The pattern holds in each row: drafting is automatic, the decision is not.
Four examples, step by step
1. A client reply
The AI drafts a reply to a client asking whether the launch can move forward a week. Before sending, the person checks three things: does the draft promise a date, does it match what the team actually agreed, and is the tone right for this client? They change "we can do that" to "we can do that if the copy is approved by Thursday", then send it themselves.
The approval point: the send button. The AI never sends.
2. Actions from a meeting
The AI reads the notes and proposes six actions with owners. The person removes one that was only discussed, not agreed, and changes an owner who was not in the meeting. The track commitments guide explains why a suggestion and a promise need to stay separate.
The approval point: before actions are sent to anyone or added to their lists.
3. A research summary with sources
The AI answers a question across ten documents and cites the passage behind each claim. The person opens two citations at random and reads them in context before using the answer in a report. If either does not say what the summary claims, the whole answer gets checked. See how to check an AI answer for a fuller routine. If you are choosing a tool for that set-up, a second brain app that answers with sources is what to look for.
The approval point: before the claim leaves your desk.
4. A spreadsheet formula
The AI writes a formula to flag overdue invoices. The person tests it on five rows where they already know the answer: one paid, one due today, one overdue, one with no date, one with a credit. Only after all five come out right does the formula go down the whole column.
The approval point: a small test before it touches the full data.
Human in the loop, on the loop, or out of the loop?
Not every task needs a person on every step. The EU High-Level Expert Group's Ethics Guidelines for Trustworthy AI (opens in a new tab) separate human in the loop (intervention in every decision cycle), human on the loop (monitoring, with the ability to step in) and human in command (deciding whether and how to use the system at all).
A practical way to choose:
- In the loop for anything that leaves your organisation, costs money, deletes something, or affects a person's job or money.
- On the loop for routine, reversible work, such as tidying formatting or filing notes, where you review a log afterwards and can undo.
- Out of the loop only for work so low-risk you would not notice a mistake, and would not mind if you did.
Human in the loop in Plexii
Plexii is a desktop workspace where each task gets its own desk, and its assistant is built around this pattern. When you ask Plexii to change something, it answers with cards: one per change, each saying what it would do and why, with a preview of the real result, such as the actual columns of a table. Nothing changes until you apply it, and you can apply one card, a selection, or all of them.
How much Plexii may do on its own is set in assistant autonomy. New accounts start on Ask before acting. Suggest only never applies anything for you, and Act on low-risk, ask for the rest applies routine changes and still waits for you on anything risky. Removing something always waits for you, at every level. In a shared organisation, admins can set a ceiling no member's assistant can exceed. The AI that asks first page shows it in one example.
Plexii's AI runs on Plexii credits or your own Anthropic key; pricing explains what is included.
Put your own approval step in place this week
- List three AI tasks you or your team already do.
- For each, write the approval point: the exact moment a person reads the real output.
- Name the risk if the step is skipped, using the table above as a guide.
- Move anything that leaves the building to "in the loop", even if it feels slow.
- Review after two weeks. If people approve without reading, make the preview smaller or the batches shorter.
More on checking AI output is in the AI you can check hub. If you want an assistant that proposes changes as cards you approve, download Plexii free.
Common questions
What is an example of human in the loop? An AI drafts a reply to a customer and a person reads, edits and sends it. The AI never sends on its own. The same shape applies to meeting actions, research summaries and spreadsheet formulas.
What does human in the loop mean in AI? It means a person reviews and decides at a set point before the AI's output has an effect. The AI prepares; the person approves, edits or rejects.
What is the difference between human in the loop and human on the loop? In the loop, a person approves each decision before it happens. On the loop, the system acts and a person monitors and can step in. In the loop suits higher-risk work; on the loop suits routine, reversible work.
Does human in the loop slow work down? A little, at the approval step. It saves time later by catching wrong promises, wrong owners and wrong figures before they spread. Keeping each batch small makes review quick.
What is automation bias? It is the tendency to rely too much on what an automated system produces. The EU AI Act names it as something people overseeing high-risk AI need to stay aware of. The practical defence is to check the real output, not a summary of it.
Can Plexii change my work without asking? Not by default. New accounts start on Ask before acting, so changes arrive as cards you apply. If you choose Act on low-risk, ask for the rest, routine changes apply themselves and risky ones still wait. Deletions always wait for you.