Human in the loop, in business process automation, means an automated system hands part of its work to a person before, during or after it acts. The phrase is older than machine learning, from control engineering and simulation, and in machine learning it usually means people labeling training data, but in a finance or operations process it describes something more concrete: who approves a payment before it goes out, who resolves the invoice the system couldn't match, and who checks a sample of the entries the system posted on its own. The volume of that work is large. Ardent Partners, in its State of ePayables 2025 research, found that 18.4 percent of supplier invoices are exceptions, so in accounts payable alone roughly one invoice in five needs a person to resolve it before it can be paid. Where that person sits, and on which actions, decides both what an automation costs and which errors reach the ledger.
What human in the loop means in business process automation
A business process is a chain of actions on records: read an invoice, match it to an order, code it to an account, approve it, pay it, post it. An automated system can take any of those actions. Human in the loop is the design decision about which ones it takes alone and which ones involve a person, and in what way. The decision is made for each action separately. In a well-run accounts payable automation, the system may match and code thousands of invoices without anyone looking, route a few hundred to a clerk, and wait for a controller's approval on every payment run.
The distinction from the machine learning sense matters because the two get confused in vendor material. A model trained with people labeling its data still needs a decision about where people sit once it runs in production. Labeling improves the model. The production design decides what happens when the model is wrong on a live invoice.
The three places a person can sit
The first place is before the action. The system prepares the work and stops, and a person approves, edits or rejects it before anything changes in the system of record. A payment run that waits for a controller's sign-off is the common case. This position catches errors before they take effect, at the price of a person's time on every case and a delay while the work waits.
The second place is on the exceptions. The system completes the cases it can decide within the rules it has been given and routes the rest to a person: the invoice with no purchase order, the price outside tolerance, the supplier name that matches two records. The person sees only the cases the system flagged. This is where most of the human work in a production automation sits, and its cost depends on how many cases reach the queue and how much evidence arrives with each one, which is the arithmetic set out in the unit economics of the exception tail. It catches the errors the system knows it might make. It can't catch an error the system made with confidence, because that case never reaches the queue.
The third place is after the action. The system acts on its own, and a person reviews a sample of the finished work: fifty posted invoices a week, every entry above a set amount, every case a new supplier touched. Sampling catches the confident errors the exception queue misses, and it measures the system's real error rate on live work, which is the number that justifies giving it more autonomy or less. It doesn't stop the error that was sampled. That entry has already posted, so sampling only belongs on actions that can be corrected after the fact.
Human on the loop vs human in the loop
The terms human in the loop and human on the loop describe two levels of supervision. A human in the loop takes part in each action before it takes effect, so the system can't proceed without that person. A human on the loop supervises a system that acts on its own: the person watches its output, gets alerted when something looks wrong, and can override a decision or stop the system. The first position above is in the loop. The third is on the loop. The exception queue sits between them, since the person is in the loop for the flagged cases and absent for the rest.
The EU AI Act writes the on-the-loop powers into law for the systems it classes as high-risk. Article 14 of Regulation (EU) 2024/1689 requires that the people assigned to oversee such a system can disregard, override or reverse its output, and can interrupt it through a stop button or a similar procedure that brings it to a halt in a safe state. Most finance and operations automations won't fall in the Act's high-risk categories, but the list is a useful specification. An on-the-loop design that gives the supervisor no way to reverse an action or halt the system isn't supervision.
The same article names the main weakness of both positions. It requires that the overseer stay aware of automation bias, the tendency to rely too much on a system's output. The evidence on that tendency is strong. Goddard, Roudsari and Wyatt, in a systematic review of 74 studies of decision support in healthcare and other fields, found that when the system's advice was wrong, it raised the risk of an incorrect decision by 26 percent across the four healthcare studies they could pool, and that in four studies people changed a correct answer to an incorrect one after seeing the system's advice in 6 to 11 percent of cases. A controller who approves a hundred correct payment runs in a row is less likely to catch the one that's wrong.
Deciding which actions need which, by reversibility and the cost of an error
Two properties of an action decide where the person should sit. The first is reversibility: whether a wrong action can be undone, at what cost and within what time. A miscoded invoice can be reclassified with a journal entry before the books close. A payment sent to a fraudster's account usually can't be recalled. The second is the cost of a single error: the loss, the rework and the exposure one wrong action creates. A wrong cost center on a $40 office supply invoice costs minutes. A wrong vendor bank account costs the whole payment.
Crossed, the two properties give a working rule. Actions that are hard to reverse and expensive when wrong get a person before the action: releasing payments, changing bank details, issuing credits above a threshold, closing a period. Actions that are cheap to reverse and cheap when wrong run on their own, with a person sampling the output after the fact: coding, matching within tolerance, routine accruals. Actions that are expensive when wrong but easy to reverse, such as a large journal entry the controller reviews before close, can run with a person on the loop and a tight sample. Actions where the system can't tell whether it is right go to the exception queue, whatever their value.
The rule works best with numbers in it. A threshold by amount, by supplier age, or by the system's own confidence turns the decision into a setting the finance team owns and changes as the measured error rate moves. We describe how those thresholds sit alongside an action log and scoped permissions in controls for autonomous agent actions.
What keeps each position working
Each position fails in its own way, and the failures are predictable. Lisanne Bainbridge, in her 1983 paper "Ironies of Automation", observed that automating the routine part of a task leaves people the hardest residual cases while taking away the routine practice that kept their judgment sharp. An exception queue with no context has that problem: the clerk gets the irregular invoices with none of the history that would explain them. The remedy is to send each exception with its evidence attached (the order, the receipt, the supplier's emails, the system's proposed resolution) so the person decides rather than investigates.
Approval before the action fails through volume. A person who approves every case approves most of them without reading, which is automation bias in its plainest form. Approval belongs on the small set of actions where it matters, with the system showing what changed since the last run and why this case differs from the routine ones. Sampling fails when nobody acts on what it finds. A sample is useful only when its error rate is recorded, compared with the target, and used to widen or narrow the system's autonomy.
The share of work each person sees should fall over time. When a clerk resolves the same kind of exception for the third time, the resolution belongs in the system's rules, and the next one shouldn't reach the queue. Closing that gap, from a system that handles most cases to one that handles nearly all of them, is the hard part of any production automation, and it's the subject of the 80-to-99% problem.
A reasonable counter, answered
A reasonable counter is that putting a person on every consequential action is the safe default, and that a finance team should approve everything until the system has proved itself. There's sense in starting there, since a new automation has no measured error rate and an approval step is how one gets measured. The problem is that it doesn't stay safe. A person approving every case at volume stops reading closely, and the Goddard review found that heavier workloads raise reliance on automated advice, and that errors rise when a system is highly but not perfectly reliable. Approval on everything also keeps the cost of the process close to where it started, so the automation never pays for itself. The safer design puts approval on the few actions that are hard to reverse and expensive when wrong, puts sampling on the rest, and uses the sample's measured error rate to decide when a threshold moves.
Common questions
- What does human in the loop mean?
- Human in the loop means an automated system can't complete certain actions without a person taking part. In machine learning the phrase often refers to people labeling training data, but in business process automation it refers to a person's role in the live process: approving an action before it happens, handling the cases the system routes out, or checking a sample of finished work. Each position catches different errors at a different cost. Most production automations use all three, on different actions.
- What is the difference between human on the loop and human in the loop?
- A human in the loop has to act before the system's action takes effect, so the system waits for that person. A human on the loop supervises a system that acts on its own, watches what it does, and can intervene, override or stop it. In the loop suits actions that can't be undone or are expensive to get wrong. On the loop suits high-volume actions that are cheap to reverse, where waiting for a person on every case would cost more than the occasional error.
- Which actions in a finance process should always need a person's approval?
- Actions that move money out of the company or change who can receive it are the usual candidates: releasing a payment, changing a supplier's bank details, issuing a refund or credit above a set amount. These are hard to reverse once they leave, and the loss from one wrong action can be large. Posting an accrual or coding an invoice to a general ledger account is easier to correct, so it can run with sampling and a person on the loop. The threshold for each action should be a number the finance team sets and reviews.