"What is human in the loop?" Owners ask it after a vendor promises it in a demonstration, or after a manager uses it to explain why a tool is safe. The honest answer is plain. Human in the loop means a named person reads, checks or approves what a tool produced before it counts. The tool drafts; the person decides. It sounds like a feature. It is a staffing decision, and it is yours.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. The failures I remember were never a tool doing something strange. They were a tool doing exactly what it was set up to do while the person who was meant to be watching had stopped looking. I am a lawyer by training, and nothing here is legal advice. Now, at 013 Labs, I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies, and the loop is the part of every plan that decides whether the rest of it is real.
The short answer
A tool produces something: a draft invoice, an answer to a customer, a schedule, a summary of a contract, a list of who is late. Human in the loop means that output waits for a named person before it becomes real. The person reads it, changes it or sends it back, and only then does it go out, get paid, get posted or get filed as true.
Three things have to be true for the loop to exist, and vendors tend to mention only the first. The person has to be named, not "someone in the office". The person has to have the authority to say no, to the tool and to the colleague who wanted it sent. And the person has to have the time, because a check that takes longer than the work it checks is a check nobody does for long.
Where the person sits is the owner's decision. Nobody else in the company has the standing to say which steps need a human and which do not, and a vendor's default setting is not a decision, it is a sales choice.
Where the loop belongs, and where it does not
The loop belongs wherever a step is a promise, a payment, a price, a hire, a deletion or a message under the company's name. Those are the steps where a mistake costs more than the minute it takes to check: a delivery date the crew cannot keep, a credit nobody approved, a reminder sent to a customer who paid last week.
The loop does not belong where the step is boring, reversible and already checked downstream. A tool that sorts supplier invoices into folders does not need a person to approve each move, because the person who pays the invoice will look at it anyway. A draft of a weekly summary needs no approval before it exists, only before it is sent. Putting a human on every step is how companies conclude that the tools are not worth it, when the loop was simply in the wrong place.
The test is simple. Ask what happens if this step is wrong and nobody notices for a day. If the answer is "a customer, a supplier, an employee or the bank acts on it", the loop goes here. If the answer is "someone fixes it in a minute when they see it", the loop goes later, or not at all. Can AI do my invoicing walks through one process with the loop placed step by step; the same shape works for almost everything else.
The three ways the loop fails
The rubber stamp. The tool produces sixty items a day, and the person approves sixty items a day, and after a week the approving takes four seconds each. The loop still exists on paper. In practice nothing is read. You find out when a customer forwards the email that should never have gone. The cure is fewer items in the loop and a check that actually requires a judgement, not a click.
The hidden loop. Everyone agrees a person checks the output. Nobody can say who. The vendor assumed your team would decide; your team assumed the vendor's tool did it; the manager who set it up left in the spring. When something goes wrong, the loop turns out to have been a sentence in a proposal. The cure is a name, written on the page where the process is described, and a log that shows that person's checks. A quieter cousin is the loop that checks the wrong thing: the person confirms the email reads well and never notices it went to the wrong customer, because the check was built around tone, not facts. Write down what the person is checking for.
The loop nobody can leave. One person approves everything, and when that person is on holiday the invoices stop, the schedule stops and the reminders stop, so somebody switches the check off "just for the week". The switch rarely goes back on. The cure is a second named person, and a rule for when both are away: the work waits, or a narrower set of steps goes through.
The data rule
The loop is also where your data rule gets enforced or ignored. The person in the loop sees what the tool was given and what it produced, so they are the one who notices that a customer list was pasted into a free tool, or that a draft quotes a price from a document that should never have been uploaded. Decide, in writing, what may be put into a tool at all: no customer lists in full, no pricing, no payroll, no bank details, nothing about an employee's health or pay, nothing a customer told you in confidence, and nothing under a non-disclosure agreement. Run every tool on a company account with written terms about who keeps the data and for how long, never on a free personal login. What is shadow AI covers the tools already in use that nobody has written down, which is where most data-rule problems start.
Five checks before you trust a loop
First, write the process on one page. Every step in order, with a mark against each step that is a promise, a payment, a price, a hire, a deletion or a message under the company's name. Those marks are where the loop goes. If you cannot write the page, you cannot place the loop, and the vendor's default will place it for you.
Second, name the person at each mark. One name per step, a second name for when they are away, and a line about what they are checking for. "Someone in accounts" is not a name. Who owns AI in a company is the wider version of the same question; the short answer is the person staff already go to.
Third, decide what the tool may do without asking. The boring, reversible steps that a downstream person will see anyway. Write those down too, so that the loop is not quietly extended by someone nervous, or quietly removed by someone busy.
Fourth, insist on a log. A record of what the tool produced, who checked it, when, and what they changed. Without the log you cannot tell a real loop from a rubber stamp, and neither can your insurer, your lender or a buyer. Guardrails and safety filters explains the technical side of what a tool will and will not do on its own; the log is the business side.
Fifth, set the thirty-day read before you start. How many items were sent back, how long items waited for a check, and how many went through without being read at all. Decide now what those numbers must look like for the loop to count as working, so the answer later is a reading, not an argument.
Check yourself
Seven yes or no questions. Each no is a decision still to make.
- For each tool in use, can you name the person who checks its output before it counts?
- Is that name written on the page that describes the process, not just known in someone's head?
- Does that person have the authority to send work back, including work a colleague wanted sent?
- Is there a second name for when the first is away?
- Do you know which steps the tool may complete without asking anyone?
- Is there a log of what was checked, by whom, and what was changed?
- Have you decided what the thirty-day numbers should look like?
Mostly no? That is the normal starting point. The AI readiness scorecard asks ten questions like these across the whole company.
Starting safely
The page from the five checks is the loop, and it is also most of what a board, a lender or a buyer means when they ask whether a person stands between the software and the figures they rely on. Do I need an AI strategy is the five-decision version of the same page, if you want the wider frame first.
Start with one process and one person. For the first month, set the tool to draft only: it prepares, the person reads, and nothing leaves the building, changes a record or reaches a customer without that person. Keep two notes through the month: what the tool got wrong, and how long the checking took. If the checking takes longer than the work used to, the loop is in the wrong place; if the tool gets the same thing wrong three times, the page is missing a rule. At the end of the month, loosen the loop one step at a time, and only on the boring, reversible steps. Can AI be my assistant is the same discipline applied to the owner's own desk, where the temptation to skip the check is strongest.
Picture it going wrong
A made-up example, to show the shape. A seventy-person commercial cleaning company puts a tool on its customer email. The tool reads each message, drafts a reply, and the office manager approves each one. For two weeks it is a relief: the inbox is clear by ten every morning. By the third week the approving has become a habit, forty clicks before the first coffee. In week four a customer writes to cancel a contract after a missed service, and the tool drafts a warm reply thanking them for their feedback and confirming the cancellation, in the company's name, with the standard thirty days' notice waived because the customer asked. The office manager clicks. The owner hears about it from the customer, who is surprised how easy it was. Nothing in the tool failed. The loop had a name, a person and a click, and it had stopped being a check sometime in week three because nobody decided what the person was checking for, and nobody looked at how many went through unread.
Mistakes that repeat
- Accepting the vendor's default for where the check goes.
- Putting a person on every step, then concluding the tool is not worth it.
- A loop with a title but no name.
- No second name, so the check is switched off the first time the approver is away.
- Measuring the loop by whether approvals happen, instead of by what gets sent back.
- No log, so nobody can tell a real check from a habit.
- Checking tone when the risk is facts.
Frequently asked questions
What is human in the loop?
A named person reads, checks or approves what a tool produced before it counts: before it is sent, paid, posted or filed as true. The tool drafts and the person decides. For an owner it is a staffing decision about where that person sits, not a feature of the software.
What is the difference between human in the loop and human on the loop?
In the loop, the person approves each item before it goes. On the loop, the tool acts on its own and the person watches the log and steps in when something looks wrong. The first fits steps that touch money, customers, employees or the company's name. The second fits boring, reversible steps that a downstream person will see anyway.
Do I need a human in the loop for every AI tool?
No, and trying to have one everywhere is how companies give up on tools that would have helped. Put the person where a mistake would reach a customer, a supplier, an employee or the bank before anyone noticed. Let the tool work alone where the step is reversible and already checked later.
Who should the human in the loop be?
The person who already owns the outcome of that step: the office manager for customer email, the controller for any number a customer or lender will see, the dispatcher for the day's board. Not the vendor, not the IT provider, not "whoever is free". Name a second person for when the first is away.
How do I know the loop is working and not just a habit?
Count three things each month: how many items the person sent back, how long items waited for a check, and how many went through without being read. A loop that never sends anything back is not a check. A loop where items wait for days is in the wrong place or has the wrong person.
Does human in the loop protect us legally?
A loop is a control, not a shield, and this is not legal advice. What it gives you is a record of who checked what, and when, which is what an insurer, a lender, a buyer or your own board will ask for. Whether it satisfies a particular rule is a question for your own adviser.
What should I ask a vendor who promises human in the loop?
Ask where in their tool the person is placed, whether that can be changed, what the log shows, how a person sends an item back, and how the whole thing is switched off. If the answers are vague, the loop is a sentence in a proposal, not part of the product.
Not sure where you stand? Take the 3-minute AI readiness scorecard.