Human in the loop means a person reviews or approves what an AI system produces before or after it takes effect. Put the person at the point where a mistake would matter most: before anything is filed, sent to a client, or paid. For low-stakes, reversible output, reviewing results afterward is often enough.
Two systems we built put a person in two very different places.
The AZLG Case Hub reads medical records for a personal injury practice and drafts a chronology. The attorney checks it before relying on it. Our Content Engine researches, writes, lays out, and publishes an article each morning. A human reviews the output, not the process.
Both are human in the loop. The difference is where the loop closes, and that choice is the most important design decision in any AI workflow. Put the person in the wrong spot and you either get a bottleneck that saves no time or an unattended system that makes mistakes nobody sees.
What does human in the loop mean?
Human in the loop is a way of building AI systems so that a person reviews, corrects, or approves the system's output at a defined point. The model does the reading or drafting, and a person who understands the subject stays responsible for the result.
The phrase covers a range. At one end, a person approves every item before anything happens. At the other, the system acts on its own and a person looks at a sample afterward. Neither end is right in general. What matters is matching the amount of review to what a mistake would cost.
A language model can produce text that reads well and is wrong. That is the reason the person is there. Review is not a courtesy to skeptics. It is part of how the system produces correct work.
Where should a person review AI output?
A person should review AI output at the point where a mistake would matter: the last step before the output is filed, sent, published under your name in a sensitive context, paid, or used to make a decision about someone. Earlier steps can run unattended if that final check exists.
To find that point, walk the process and ask at each step what happens if the output is wrong and nobody notices.
- List every step from input to final effect.
- Mark the irreversible ones. Filing, sending, charging, deleting.
- Mark who is affected. A draft seen only by staff is different from something a client or court receives.
- Place review before the first irreversible step that touches someone outside the business.
- Let everything before that run, with logs, so the reviewer can see how the draft was produced if they need to.
In the case hub, the chronology feeds legal work, so the attorney checks it. The software does the slow reading. The attorney does the part that carries responsibility.
Should a human review before or after the AI acts?
Review before the AI's output takes effect when the result is hard to undo or the stakes are high. Review afterward when the output is low risk, easy to correct, and produced in volume. Many systems use both: approval for the risky steps and after-the-fact checks for the rest.
| Situation | Review before | Review after |
|---|---|---|
| Court filings and legal documents | Yes | |
| Messages sent to clients | Yes | |
| Payments and refunds | Yes | |
| Summaries used to make decisions | Yes | |
| Routine articles on an established site | Yes | |
| Internal tags, labels, and sorting | Spot checks | |
| Drafts that staff will edit anyway | Built into the editing |
Our content engine sits in the right-hand column. Each morning it picks a topic from keyword research, writes the article, converts it to the site's layout, sets metadata, links it into related pages, and publishes. A person reviews what was published. A published article can be corrected or removed, which is what makes that arrangement reasonable. We describe the pipeline in automated content publishing.
That design would be wrong for a document going to a court.
How do you make human review fast instead of a bottleneck?
Human review stays fast when the reviewer is checking a draft against its sources, not redoing the work. Show where each statement came from, limit review to the parts the model produced, and make corrections a single action. If checking takes as long as writing, the workflow has failed.
The case hub is built around that idea: a chronology the attorney can check instead of write. Checking is faster than writing only if the draft is organized for it.
Things that help:
- Sources beside claims. Each entry points to the record it came from, so the reviewer can confirm it without searching.
- Only the variable parts. Fixed text assembled by code does not need rereading. We explain that split in when to use AI and when plain code is better.
- Easy correction. Edit in place, approve, move on.
- A record of changes. What the reviewer fixed shows where the system needs work.
What goes wrong when there is no human in the loop?
Without a person in the loop, errors pass through at the speed of the software. A model that misreads a document or invents a detail will do so fluently, and with no reviewer the mistake reaches a client, a court, or the public before anyone knows it exists.
There is a subtler failure too: review that exists on paper only. If a person is asked to approve hundreds of items a day with one click, they will stop reading. That is a rubber stamp, and it gives false comfort. A review step has to be sized so a person can do it properly.
The opposite mistake is also common. Requiring approval on every trivial step makes the tool slower than doing the job by hand, and people stop using it.
Frequently asked questions
Does human review cancel out the time AI saves?
Not when it is designed well. Reading and correcting a good draft takes less time than producing it from nothing. The saving disappears only when the draft is poor or when the reviewer has to hunt for the sources themselves.
Who should the human reviewer be?
Someone who could have done the task without the software and is accountable for the result. For a legal chronology that is the attorney. A reviewer who cannot tell a right answer from a plausible one adds no protection.
Can the amount of review be reduced over time?
Sometimes, for low-risk output with a long record of few corrections. Reduce it by sampling, not by removing it, and keep full review on anything that is hard to undo.
Is human in the loop required by law?
Rules differ by industry and location, and some professions have their own guidance on supervising automated work. Check the rules that apply to you or ask a professional who knows them.
Can one AI check another instead of a person?
A second automated check can catch some errors and is worth having. It does not carry responsibility, so it supplements a person and does not replace one.
If you are designing an AI workflow and want a second opinion on where the review belongs, walk us through the steps.