Human-in-the-loop AI explained

Human-in-the-loop AI is a setup in which a person reviews, approves or corrects what an AI system proposes or is about to do, at predetermined moments. The AI does the searching, calculating and preparation; the person takes the decisions that have consequences for customers, money or agreements. It only works if the person gets enough information to genuinely judge, and not so many requests that they start approving blindly.

What forms of human oversight are there?

There are broadly three ways to involve a person in AI.

Human-in-the-loop. The person is inside the process. Without their approval, the action does not happen. Example: the AI prepares a corrective invoice, an employee approves it, and only then is it sent.

Human-on-the-loop. The AI executes, the person supervises and can intervene. Example: payment reminders go out automatically, but an employee sees a daily overview and can stop a sequence.

Human-in-command. The person sets the boundaries: which tasks AI may do, with which permissions and up to what limit. That always applies, including with the other two forms.

For revenue tasks a combination is usual: in-the-loop for everything external or irreversible, on-the-loop for routine work with small, recoverable consequences. The distinction is closely tied to the difference between AI recommendations and AI decisions.

Why keep a human in the loop?

It is tempting to see human-in-the-loop as a temporary measure until the AI is good enough. That is a misconception. There are reasons that remain, however good the model becomes.

Information outside the system. The account manager knows the customer was promised a discount by phone last week because of a delivery problem. It is not recorded anywhere. The AI sees an invoice that is too low; the person knows why.

Commercial judgement. A justified corrective invoice of EUR 400 to a customer who is weighing up a large renewal may be better held back. That is not a calculation but a judgement call.

Accountability. Someone has to be able to explain why a decision was taken, to a customer, an auditor or a board. An AI cannot carry that.

Learning. Every correction by a person is information about where the system goes wrong. Without a person in the loop, that feedback disappears.

Where do you place the checkpoints?

The design is about placement, not quantity. Too many checkpoints make AI useless; too few make it risky.

A practical classification for revenue processes:

The thresholds in this table are examples. How to measure confidence in order to choose such a threshold is covered in how you know whether an AI recommendation is reliable.

The real risk: rubber-stamping

Human-in-the-loop rarely fails because there is no person. It fails because the person is there formally but not in substance. Someone who receives a hundred requests a day, ninety of which are right, starts clicking without reading. This is called automation bias, and it turns the check into a formality.

How to prevent it:

• Fewer, more important requests. Let internal, reversible actions through without approval and only ask for approval where it matters.
• Evidence attached. Every request shows the source, the calculation and the difference. Reviewing becomes looking rather than searching. See why AI output has to be checked.
• Ranking. Highest impact and lowest confidence at the top.
• Rejecting must be as easy as approving, and asks for a short reason.
• Check the checker. Periodically look at a sample of approved actions. If they were often wrong, the loop is not effective.

Worked example: what a well-designed loop delivers

Worked example: suppose your invoicing team of three manually checks 400 contract invoices against the contracts each month. That takes about 60 hours, and some errors still slip through because the work is monotonous.

An AI system compares all 400 automatically. It flags 35 discrepancies, each with source and calculation. The team reviews those 35, five minutes each on average: just under three hours. Say 25 are valid, with an average of EUR 450 in missed revenue a month.

• Found per month: 25 x EUR 450 = EUR 11,250.
• Time: from about 60 hours to about 3 hours plus sampling.
• The 10 invalid findings never reach the customer.

The person stays in the loop but only does the part where human judgement adds something. The figures are assumptions; the pattern is general. The type of leak in this example is worked out in how to check that contracts are billed correctly.

The loop works both ways

A person in the loop is not just a gatekeeper. Every review is also information. Someone who rejects a finding usually knows why: the discount was agreed verbally, the contract was changed last month, the customer has a dispute in progress. If that reason is recorded, the system can learn from it or someone can tackle the cause.

Three kinds of feedback that deliver the most in revenue processes:

• Missing agreement. The finding was invalid because an agreement was made outside the system. Fix: record the agreement after all, so the next check sees it. Often this is itself a process problem that costs revenue even without AI.
• Wrong source data. The finding was invalid because a field in the CRM or the ledger was wrong. Fix: correct the source, not just dismiss the finding.
• Wrong reasoning. The data was right, but the conclusion was not. Fix: adjust the rule or the model.

Without that feedback, your team reviews the same invalid findings again every month. With it, the number of requests falls and the share that is valid rises. That is the sign the loop is working.

Enforcing it technically

A person in the loop has to be built into the software, not into an instruction to the model:

• Separate preparing from executing. The AI has tools to create drafts, not to send them. Sending is a separate action that only a person can start.
• Approval is recorded. Who, when, which version. That is also your audit trail.
• What happens without a response? A request that is not reviewed within a set time expires or escalates. It is never executed silently.
• An emergency brake. Someone can pause an agent or automation in a single action.

How permissions and tools set that boundary is covered in what is an AI agent.

How do you design a human-in-the-loop process? Step by step

1. Map the process and mark every step AI can take over.
2. Mark the steps with external, financial or irreversible consequences. That is where an approval point goes.
3. Decide who approves, per type of action, with a deputy.
4. Design the request: what does the reviewer see, and can they judge it in a minute?
5. Set thresholds for amount and confidence.
6. Measure how many requests per week, how many rejected, how many approved actions later proved wrong.
7. Adjust. Actions that were always valid for months can move to on-the-loop. Actions that often go wrong get an extra check.

In a revenue intelligence system

In revenue control, human-in-the-loop is not a brake on AI but what makes it usable. The AI finds the discrepancies in thousands of records that people miss. The person decides what happens with them. RiOS works on that principle: automations only run after approval, and agents work within the rules the business sets, with human approval wherever the business wants it. How this fits into the whole AI setup is explained in how AI works within Revenue Intelligence.

Frequently asked questions

Does human-in-the-loop not make AI slow?

Only if the loop is badly designed. If approval is limited to actions with consequences and the evidence is attached, it costs minutes a day. The AI does the time-consuming part.

Who should the human in the loop be?

Someone with knowledge of the customer or the process and the authority to decide. For a corrective invoice that is often invoicing or the account manager; for a rate change, someone with commercial responsibility.

Can the human in the loop ever be removed?

For internal, reversible and predictable actions, it can move to oversight afterwards. For actions towards customers or with financial consequences, a person beforehand is the sensible default.

Is human-in-the-loop a legal requirement?

For certain applications there are legal requirements for human oversight, for example for fully automated decisions about individuals. For most B2B revenue tasks it is above all a sensible design choice. Check the rules in your jurisdiction, and take legal advice if you are unsure about your own situation.