Knowledge · human in the loop

Human in the loop: AI with a person who approves

Human in the loop means that AI does the work and a person checks it before it goes to a customer or ends up in a system. That's how you use AI safely in your business processes, and for most tasks it's simply sensible too. Here you'll read about the four roles between human and AI, what approval looks like in practice without it becoming extra work, and when AI may do it entirely on its own.

The four roles Approval without extra work

By Olaf Lemmens
Updated 27 October 2026
10-minute read

A day with human approves

Customer service

08:00Drafts are readyThe AI read the new questions and prepared an answer for each one, with the order information included.
08:30Read, adjust, sendApprove with one click, or change a sentence first.
10:45ExceptionA complaint about damage. The AI is in doubt and passes it to the team leader.
14:00Spot checkFor what the AI already handles itself, someone checks a handful of cases.
16:30FeedbackWhat was often changed? That goes into the instruction, for tomorrow.

What is human in the loop?

Human in the loop means there's a person in the process who checks the AI's work before it goes out the door. The AI reads, looks things up and prepares. The employee checks it, adjusts where needed and approves. Only then does the email go to the customer or the order go into the ERP.

In Dutch it's also called 'mens in de lus'. In the APO method we call this role 'human approves'. It's one of the four roles between human and AI, and by far most opportunities land here or on 'AI assists'.

Compare it with a new colleague who works very fast. You let them do the groundwork, but you read their first quotes before they're sent. The difference: this colleague has already prepared all the new questions by eight in the morning.

That's how it works at our clients too:

At an enterprise client, 80 percent of the 6,000 customer questions a month have a prepared answer before the ticket opens, compared with 30 percent at first. An employee checks and sends it, always.
At Van Berkel Professionals, the incoming stream of requests is read, filtered and entered into their own systems. A person approves every request. That saves 192 hours a month.
At KplusV, AI assesses pitch decks against their own canvas and finds suitable funders. A consultant always checks the result.

The AI does the work. The human holds the pen.

The four roles between human and AI

AI isn't all or nothing. Between 'the human does everything' and 'the AI does everything' there are two roles, and that's where most tasks land. Which role fits depends on how suitable the task is for AI and on three questions: is creativity the core, is physical presence needed, and does a person carry personal or legal liability?

1. Human

The human does the work, AI stays out. That applies when the task isn't very suitable for AI, or when creativity or physical presence is the core. A conversation with an angry customer, a technician dealing with a breakdown, coming up with a new campaign.

2. AI assists

The human does the work and decides, AI does the groundwork: a first draft, a summary, looking something up. This is the role where personal or legal liability is involved. Think of an advisory report the expert puts their name to, or the decision to reject an applicant. AI may draft the letter, the decision stays with a person.

3. Human approves

AI does the work, a person checks it before it goes out. This is human in the loop, and the role for most tasks AI can take over. Answering customer questions, drafting quotes, writing product copy, processing requests.

4. Fully AI

AI does it itself, a person spot-checks and monitors it afterwards. Only when the task is very suitable for AI and no creativity, physical presence or liability is involved. Think of reading and booking invoices with a fixed format, with a check on the purchase order number.

In the APO method, the role per process step follows from an AI fit check of seven questions, plus these three. Try it yourself below with the role meter, or do the check for a task from your own work on the page about the APO method.

What does 'human approves' look like in practice?

Usually in one of three ways: the AI prepares a draft that a person reads and sends, the AI handles the ordinary work and sends exceptions to a person, or the AI works on its own and a person checks samples. Often you combine them, and shift along as trust grows.

Draft prepared, person sends

The AI reads the question, looks up the customer and the order and prepares an answer in the helpdesk or as a draft in the mailbox. The employee reads it, changes a sentence if needed and presses send. In your ERP it works the same way: a draft order or draft journal entry that only becomes final when someone confirms it.

This is the default to start with. Nothing goes out without a person having seen it, and you quickly learn what the AI is good at and what it isn't.

Exceptions to a person

The AI handles or prepares the ordinary work, and sends anything that deviates to a person. You define beforehand what counts as an exception:

the AI is in doubt, or is missing information
the amount is above a threshold
the customer is angry, or wants to cancel
it's a type of question the AI hasn't seen before
it's about a complaint, damage or something legal

That way the employee spends their time on the cases where they really make the difference.

Spot checks

For tasks the AI does on its own, a person regularly checks a sample. A lot at first, less later, as long as it goes well. If the reviewer finds a mistake, the sample goes back up and the instruction is adjusted.

And with an AI agent?

With an AI agent that takes steps in your systems itself, you define per action what it may do on its own and what needs approval. It may look things up itself. Prepare a draft too. Sending, changing or paying for something goes past a person. How an agent works is explained in what is an AI agent?

How do you keep approval light?

By designing the check as carefully as the AI itself. If checking takes as long as doing the work yourself, you've gained nothing. Designed well, approval takes seconds per case: everything is on one screen, the source is included, and the AI itself says where it's in doubt.

This is how you do it:

Everything on one screen. The original question, the draft and the data it's based on. Nobody has to open three systems to see whether it's right.
The source included. With an answer, the passage from the policy; with an order, the lines from the PDF. Checking becomes pointing instead of searching.
Flag what deviates. Have the AI indicate where it's in doubt or what's different from normal, such as a new item or an unusual amount. Then the reviewer knows where to look.
One click. Approve, adjust and reject in the screen where the employee already works, such as the helpdesk or the mailbox. No extra tool.
Batch the work. Checking a row of drafts one after another is faster than constantly switching in between.
Learn from adjustments. Keep track of what gets changed and why. If the same correction keeps coming back, adjust the instruction. Then nobody has to fix it again tomorrow.

Do watch out for blind approval. If everything always looks fine, sooner or later someone starts clicking 'ok, ok, ok'. So vary the checks, have a second person look along now and then, and check in your log how long an approval takes. Two seconds for a quote is no longer a check.

When can fully AI work?

When four things are true at the same time: the task is very suitable for AI, no creativity, physical presence or liability is involved, a mistake is cheap to fix, and you've already shown that it goes well. Then the AI may do it itself, with checks in the workflow, monitoring afterwards and spot checks.

Characteristics of a task that can be done fully by AI:

A fixed format and a stable pattern. The same kind of input every time, the same kind of output.
Lots of examples of good work. So you can test and measure how often it goes well.
A short chain of reasoning. A few steps, no long deliberation.
A mistake stands out and is cheap to fix. For example because there's a check against another system.
No explanation needed. Nobody has to justify afterwards, legally or on substance, why it went that way.

A good example is reading and booking invoices with a fixed format, with an automatic check on the purchase order number. Or sorting and labeling incoming email for internal use. Not good candidates: anything that lands with a customer without a check and has consequences, decisions about people, and money above a threshold.

Even with fully AI, a person stays involved, just not in every case. Someone monitors the numbers, checks samples and is the owner. That's why the APO method never calculates with 100 percent saving, but with 90 percent per step at most.

From 'human approves' to more room

Start strict and give room based on numbers, not gut feeling. That's how you build trust, in yourself and in the team.

01Start with 'human approves' for everything. Every draft is checked.
02Measure how often it's approved unchanged. Per type of task or question, not in one number for everything.
03Improve the instructions based on what gets changed.
04Give room per category. If a type of question goes well for weeks on end and a mistake is cheap, it can move to spot checks. The rest stays with the human.
05Keep the way back open. If something changes, such as a new model, new policy or a new product line, go back to 'human approves' for that category until the numbers are right again.

That's also how you bring the team along. Employees who see for themselves how often the AI gets it right, and who get a say in when something needs less checking, work with it instead of around it.

Using AI safely: why we always keep a human in the loop

Because AI makes mistakes, and a mistake a person catches in time costs nothing. A language model can convincingly make something up; see AI hallucinations. In a chat you notice that yourself. In a process it ends up with a customer or in your books, unless someone sits in between.

Using AI safely takes more than a human in the loop. This is what we arrange for every process:

A human in the loop, with clear agreements about who approves what
The smallest permissions in your systems: the AI may only do what's necessary
A log of what the AI did and who approved it
Data in the EU, on your own infrastructure or with a European provider, and a way of working in line with ISO 27001 and NEN 7510 (certification is in preparation)
No lock-in: code, data and documentation are yours

And the law? Whether human oversight is mandatory depends on the application. For some applications, such as AI in recruitment and selection, the AI Act sets requirements for human oversight. When in doubt, a lawyer has the final say.

We only build agents and workflows within the AI Partnership, from € 4,500 per month, after the € 6,000 AI Quick Scan. In that Quick Scan we decide per process step which role human and AI get. All amounts exclude VAT. How that works out in customer service is on AI by department.

Try it yourself

Where does your task land?

Pick a task and see which of the four roles it usually lands on, with the reason. You'll spot the pattern quickly: most tasks land in the middle, on 'AI assists' or 'human approves'.

Pick a task

Human
A person does the work. AI stays out.
AI assists
A person does it, AI drafts or looks things up.
Human approves
AI does the work, a person checks before it goes out.
Fully AI
AI does it by itself. A person spot-checks.

Where it lands: Human approves

AI prepares the answer with the right order details. An employee reads along and sends: tone and exceptions stay human.

In one table

The four roles side by side

Who does the work, who decides and where is the check? You decide the role per process step, not per process.

	Human	AI assists	Human approves	Fully AI
Who does the work	The human	The human, with groundwork from AI	AI	AI
Who decides	The human	The human	The human approves	AI, within fixed rules
Check	Not applicable	The human chooses what's usable	Every case, before it goes out	Spot checks and monitoring afterwards
Fits when	Creativity or physical presence is the core	A person is personally or legally liable	The task suits AI, but a mistake has consequences	The task is very suitable and a mistake is cheap to fix
Example	Conversation with an angry customer	Advisory report, rejecting an applicant	Customer question, quote, product copy	Reading and booking invoices with a fixed format
How often	Where AI adds little	Often	Most often	The exception

Swipe sideways to see all columns →

Within one process, one step can be fully AI and the next 'human approves'. That's why we look per step.

Questions

Frequently asked questions about human in the loop

What does human in the loop mean?

Human in the loop means that AI does the work and a person checks it before the result goes to a customer or ends up in a system. The AI reads, looks things up and prepares a draft; the employee checks it, adjusts where needed and approves. In the APO method this role is called 'human approves', and most tasks AI can take over land here.

What's the difference between 'AI assists' and 'human approves'?

With 'AI assists', the human does the work and makes the decision; AI only does groundwork, such as a draft or a summary. That fits tasks with personal or legal liability, such as an advisory report. With 'human approves', AI does the work and the human checks the result before it goes out, as with an answer to a customer question or a draft order.

Doesn't AI with human oversight just make things slower?

Not if you design the check well. If the question, the draft and the source are on one screen and you can approve with one click, checking takes seconds instead of minutes. Employees read and adjust instead of searching and typing. That's how, at one of our enterprise clients, 80 percent of customer questions have a prepared answer before the ticket opens, and an employee checks every answer.

When can AI work without human oversight?

Only when the task is very suitable for AI, no creativity, physical presence or liability is involved, a mistake is cheap to fix and you've already shown that it goes well. Think of reading and booking invoices with a fixed format, with a check on the purchase order number. Even then a person stays involved through spot checks and monitoring afterwards.

How do you use AI safely in your business?

Start with a person who approves before anything goes out or ends up in a system. Give the AI only the permissions it needs for the task, keep a log of what it does, and make sure the data stays in the EU with a data processing agreement. Only give the AI more room when the numbers show it's going well, and keep the way back open.

Is human in the loop mandatory?

That depends on the application. For some applications, such as AI in recruitment and selection, the AI Act sets requirements for human oversight. When in doubt, a lawyer has the final say. NinA always keeps a human in the loop in any case, even where it isn't mandatory.

How do you stop employees from approving everything blindly?

Make the check meaningful and keep it sharp. Have the AI flag where it's in doubt or what deviates, so the reviewer knows where to look. Vary the checks, have a second person look along now and then and check in the log how long an approval takes. And feed adjustments back into the instructions, so the reviewer sees that their work makes a difference.

Can an AI agent also work with a human in the loop?

Yes, and that's how we build them. You define per action what the agent may do on its own and what needs approval. Looking things up and preparing a draft it may do itself; sending, changing or paying goes past a person. That way you get the speed of an agent without a mistake reaching a customer or your books unseen. More in what is an AI agent?

Where do you need a human in the loop?

Book a fifteen-minute intro call. We look at one process and decide together, step by step, what AI can do and where a person approves.

01
Intro call
Fifteen minutes, no strings. Where does work get stuck and what do you want to achieve?
02
AI Quick Scan
One day on site. Within a week, a business case and a roadmap.
03
AI Partnership
A dedicated team that builds, trains and shows what it delivers every month.

Book an intro call Prefer WhatsApp Or send a message

160+ organizations · Sessions rated 9.3 · Based in Amsterdam