Pick the control by what a wrong output costs and whether you can undo it. Includes a six-question test, eight sorted workflows and a monitor spec.
Choose the control by consequence, reversibility, volume and how fast you would notice an error, not by habit or speed. On the loop needs more design, not less: a named monitor, a sample, alerts and stop conditions. Use the test, check the rules where you operate, and move a workflow between models only as the sample record earns it.
Teams tend to pick between approving every AI output and watching a dashboard based on how busy the reviewers are, not on what a wrong output costs. Approving everything feels safe and slows down until people stop reading. Watching feels modern and can run for weeks before anyone spots a pattern of errors. Both are real controls, and each one fits some work and fits other work badly.
The words come first, because they get used loosely. This article uses them like this:
Usage varies, and plenty of business writing calls a human touch anywhere in a process “human in the loop,” including the person who wrote the prompt a month ago. That version can’t tell you who is responsible for what, so it’s not the one used here.
The terms come from the military and autonomous-systems world. A Congressional Research Service primer on US policy for autonomous weapons uses all three: human in the loop for semi-autonomous systems that engage targets a human operator has selected, human on the loop for human-supervised systems where operators can monitor and halt an engagement, and out of the loop for full autonomy.[1] The policy underneath it, Department of Defense Directive 3000.09, was first issued in 2012 and asks that such systems allow “appropriate levels of human judgment over the use of force.” As far as I could read, the directive’s own text defines autonomous and semi-autonomous systems but doesn’t use the three loop phrases.[2] So the vocabulary is borrowed shorthand, and borrowing it means deciding what you mean by it.
What follows starts after the earlier decisions are made. If you’re still deciding whether a human should be involved at all, Why Human in the Loop Matters covers the case, and the decision matrix on what to hand to AI covers what to delegate. This piece is about the next question: once AI is doing the work, which kind of control goes around it. It’s written for marketing leads, finance managers, HR business partners and support team leads, and the role differences get their own section.
The two models differ in when the person acts and what they actually look at. Everything else follows from that: the cost in time, the way each one fails, the kind of work each suits.
| In the loop | On the loop | |
|---|---|---|
| Who acts | The person approves, edits or rejects; the AI proposes | The AI acts within set limits; the person supervises |
| When the person acts | Before the output takes effect | While it runs and after, with the power to pause or reverse |
| What gets checked | Every item | A sample, anything over a threshold, anything an alert flags |
| What it costs in time | Reviewer time grows with every item produced | Design time up front, then a fixed monitoring slot |
| Typical failure | Reviewers skim as the queue grows, so approval becomes a formality | A problem slips under the sample or a threshold is set too loose, and it runs unseen until someone looks |
| Example workflows | Campaign copy before it publishes; shortlists of candidates; replies to complaints | Matching clean invoices to purchase orders; small refunds under a cap; routine FAQ replies |
A working comparison by Future Factors, not a standard. Real workflows often use both: in the loop above a threshold, on the loop below it.
The failure in the first column is the one people underrate. A marketing team puts approval on every AI-drafted social reply. In the first week the approver reads each one closely. Once the queue is long and the replies keep coming out fine, the approvals get quicker, and what’s being checked is the approver’s attention rather than the reply. NIST’s generative AI profile has a name for the pull behind this, automation bias, which it describes as excessive deference to automated systems, and it notes that it can make other problems such as made-up content worse.[3]
A person in the loop who defers is a signature, not a control.
The failure in the second column is the mirror image. No check stands between the AI and the outcome, so a bad pattern keeps running at full speed. That’s no argument against on the loop. It shows that each model has a failure you have to design against, and which one you can afford depends on the work.
The time cost runs in opposite directions too. In the loop is cheap to set up and expensive to run, because the load rises with every item. On the loop is the reverse: the setup takes real thought, and the daily cost is a slot in someone’s calendar. Teams that choose on setup effort alone usually pick in the loop and then discover the running cost around the time the queue backs up.
Neither model is safer by default.
In the loop is safer when a person can really judge each item and the volume lets them. On the loop is safer when a tired approver would be worse than a good sample, and when mistakes are cheap to catch and undo. What decides it is the work.
Four facts about the work do much of the deciding:
Regulation takes a similar line. The EU AI Act asks that human oversight measures for high-risk systems be commensurate with the risks, level of autonomy and context of use.[4] That’s proportionality, not a rule that everything needs an approver. If you operate in the EU, which of your systems count as high-risk is a question for counsel.
The test below puts the four facts, plus two that people forget, into six questions you can answer in about ten minutes for one workflow. The reading guide under the table is our working judgment, not a standard.
| Question | Points to in the loop | Points to on the loop | Points to neither (keep it human, or don’t run it) |
|---|---|---|---|
| 1. If this output is wrong and goes out, what happens? | A customer or colleague is affected and we’d have to correct it | A small, local problem someone fixes in the normal course of work | Harm to a person, a legal commitment, or a public statement we can’t retract |
| 2. Can we reverse it? | Reversible, but with effort or embarrassment | Easily reversed or corrected, ideally by a stop button or an undo | Irreversible once sent or posted |
| 3. How many a day? | A number one person can read closely | Too many to read, but patterns show up in a sample | Whatever the volume: volume isn’t what decides it |
| 4. How fast would we notice a bad one? | Slowly, so the check has to come first | Fast: complaints, an exceptions list or a reconciliation shows it within a day | We’d find out only when someone outside complains |
| 5. Does a law, contract or policy require a person to decide? | A person must be able to review before it takes effect | No such requirement, and policy allows monitoring | A rule says a named person decides, or decides about a person’s job, pay or record |
| 6. Can the reviewer see enough to judge? | Yes: inputs, sources and reasons are on the screen | Yes for a sample, and the alert tells them why it fired | No: they’d be guessing, and approving would be a formality |
Answer for one workflow at a time. Reading guide (working judgment): if the answers fall mainly in the middle column and none in the right-hand one, on the loop with a sample and stop conditions is reasonable. Mainly the left column means in the loop. One answer in the right-hand column on question 1, 2 or 5 settles it: in the loop at minimum, and often not a candidate for AI action at all. Question 3 doesn’t settle it alone.
One more thing the test shows: the same workflow can land in two places. Refunds under a set amount may sit on the loop while larger ones go in front of a person. That split by threshold is common, and it’s often a sensible design.
Choose the control by what a wrong output costs and whether you can take it back, not by how busy the reviewers are.
Here is the test applied to eight workflows a marketing, finance, HR or support team might run. They’re filled in to show the reasoning, and your answers will differ with your own volume, systems and rules.
| Workflow | Control | Why | What the person checks |
|---|---|---|---|
| Publishing campaign copy | In the loop | Public, hard to retract, brand and legal claims at stake | Every piece: claims, tone, product facts, then approve |
| Refund approvals | On the loop under a cap; in the loop above it | Small refunds are reversible and high volume; large ones are real money | Sample of capped refunds weekly; every refund over the cap before it is paid |
| Internal meeting summaries | On the loop | Low consequence, quick to correct, fast to spot | A few summaries a week against the source; action items confirmed by their owners |
| Candidate shortlisting | In the loop, with HR and legal sign-off | A decision about a person’s access to work; rules in some places add duties, so ask counsel | Whether anyone filtered out should have stayed in; the criteria themselves |
| Invoice matching | On the loop for clean matches; in the loop for exceptions | Rules are checkable and errors show up in reconciliation | The exceptions list daily; a sample of auto-matched invoices; the exception count over time |
| Social reply drafts | In the loop for complaints and sensitive topics; on the loop for routine questions | A reply is public and fast; routine ones are low risk | Complaints before posting; a sample of routine replies each day |
| Ad budget changes | On the loop with a hard cap | Spend is fast and partly reversible, so a cap and an alert limit the damage | Daily spend against cap; a change over a set size; a stop condition if cost per result jumps |
| Answers to staff on pay or policy | In the loop until the sample record is clean, then on the loop for plain lookups | A wrong answer about entitlements affects a person’s money | Sample read weekly; anything touching pay or discipline routed to a person |
Working judgment by Future Factors, to show the reasoning. The right control depends on your own volume, tooling and the rules that apply to you.
Look at where the “on the loop” rows get their safety. It comes from the cap, the exceptions list, the sample and the alert, not from the AI. A finance team that lets AI match invoices to purchase orders and post the clean matches on its own is relying on something to say when the matching has gone wrong. One morning the exceptions list is empty. That can mean everything matched, or it can mean a rule changed and no flags are being raised, and from the list alone you can’t tell which. A monitor who has only ever looked at that list won’t know either. That’s the point where on the loop stops being a control and becomes a hope.
The shortlisting row is in the loop for two reasons, consequence and rules, and neither goes away when volume is high. It’s also a place where the person in the loop can slip into approving whatever the tool ranked first, so the check needs to look at who was filtered out, not only at who was ranked high.
A support team lead sees the volume problem from the other side. The team gets a steady stream of routine questions, and approving every drafted reply would mean the lead spends the day clicking. Moving the routine ones on the loop, with a daily sample and a pause rule, gives the lead time to put real attention on the complaints, which stay in the loop.
The control goes where the attention pays off best.
On the loop takes one line to describe and much longer to build well. “A person monitors it” is not a design. It leaves out who, how often, looking at what, with what power, and what makes the whole thing stop. Those are the parts that decide whether the loop exists.
The EU AI Act’s wording for high-risk systems is a useful checklist even if the Act doesn’t apply to you. The people given oversight need to understand the system’s limits, stay aware of the pull to over-rely on its output, interpret that output, decide not to use it, override or reverse it, and intervene or interrupt it through a stop button.[4] Deployers are asked to give that oversight to people with the necessary competence, training, authority and support.[5] Authority and support are the words organisations skip. A monitor with no time in their week and no standing to pause a live process is a name on a list.
Here is the split for three on-the-loop workflows, including the part teams leave blank: what the monitor actually reviews. The numbers are starting values to adjust against your own sample record, not benchmarks.
| Workflow | What AI does | What the human still owns | Sampling, thresholds and stop conditions |
|---|---|---|---|
| Refunds under the cap | Approves and issues refunds that meet the written policy | The policy and the cap; every refund above it; the call to pause | Sample about one in ten each week, or at least twenty, whichever is more; alert when refund volume or average size jumps; pause if a sample shows repeat wrong approvals |
| Invoice matching | Matches invoices to purchase orders and posts clean matches | The matching rules; every exception; month-end sign-off | Exceptions list read daily; sample of auto-posted items weekly; alert if the exception rate drops sharply or spikes; pause on a rule change until a sample is checked |
| Routine FAQ replies | Sends replies to a list of approved question types | The list of question types; wording of the answers; anything outside the list | Daily sample of sent replies; alert on complaint or reopen; pause the reply type if the sample shows a wrong answer on pay, safety or policy |
A copy-ready starting point. Replace the numbers with ones your monitor can actually run, then adjust them with the sample record.
Notice what each row has in common. There’s a thing the AI is allowed to do, a boundary, a sample somebody reads on a schedule, an alert that doesn’t depend on someone remembering, and a stop condition written down before the first run.
If you can’t fill all four cells for a workflow, it isn’t ready for on the loop.
A support team lead reading a daily sample of replies is doing this well when they are reading to find a pattern, not to approve individual items. Ten wrong answers about one policy are one finding, and the fix is in the source document or the list of approved question types, not in ten edits.
Alerts deserve a second thought. One that fires every day gets muted, and one that stays silent for months gets forgotten. Set the threshold so it fires rarely enough to be read and often enough that the monitor has seen it work, and test it with a deliberately bad example before go-live.
On the loop means a named person with a scheduled slot, the authority to stop it, and stop conditions written down before it starts.
On the loop is the wrong choice in three situations, and it helps to know them before you’ve built the dashboard.
The EU AI Act gives one concrete example of how much stricter in the loop can get. For some remote biometric identification systems, it says no action or decision is taken on the strength of an identification unless it has been separately verified and confirmed by at least two people with the necessary competence, training and authority.[4] It also makes exceptions, including for certain law enforcement, migration, border control and asylum uses. You won’t be building that, but it shows where the dial goes when the stakes are high: more people, with more authority, before anything happens.
In the loop can be done badly too, and it’s worth being careful here. It fails when the reviewer can’t see enough to judge, when there are more items than a person can read, or when the approval step exists to give cover to a decision that was already made. If you can’t fix those, moving the same work on the loop with better limits is sometimes a more honest design than a rubber stamp. Sometimes the answer is that the AI shouldn’t do that piece of work at all, and that’s a legitimate result.
Out of the loop can be a fair choice too, for trivial, reversible work such as sorting your own files into folders. The mistake is arriving there by accident, which is how a workflow tends to end up unmonitored: someone built it quickly, it worked, and the review step didn’t get added.
For a wider view of how people and AI share a workflow, see What Is an AI-Human System, and for five checks to run before you scale a workflow, see the AI Trust Checklist.
The same two models look different from different desks, because the volume, the reversibility and the rules differ. Here is how the starting point shifts by role.
| Role | Typical starting control | Design this first |
|---|---|---|
| Marketing lead | In the loop for anything public (copy, claims, replies to complaints); on the loop for internal summaries and reporting drafts | A named approver per channel, and a list of claims that don’t get published without a source |
| Finance manager | On the loop for rule-based matching under a cap; in the loop for payments, accruals and anything that leaves the company | The exceptions list, the cap, and who signs at month end |
| HR business partner | In the loop for anything that touches a person’s job, pay or record; on the loop for policy lookups once the sample record is clean | Who reviews, what they can see, and how a person challenges a result |
| Support team lead | In the loop for complaints and edge cases; on the loop for routine questions | The list of approved question types, the daily sample and the pause rule |
Starting points by role, working judgment. Adjust for your volume, tools and the rules where you operate.
Each role also has a blind spot worth watching. Marketing leads can under-check routine content because it feels low stakes. Finance managers can trust an exceptions list without asking what it can’t see. HR business partners can accept a summary in place of the notes behind it. Support leads can read a sample for tone and miss a wrong policy detail.
Oversight is also a skill, and it’s one that teams tend to assume people already have. Reading a sample for patterns, setting a threshold, deciding when to pause: none of that comes automatically with a job title. The four things involved stay separate. Training builds the capability, so a team lead can read a sample and set a stop condition. Workflow design gives that capability a recurring task to land in, such as the Monday sample. Reinforcement, like a manager who asks what the sample found, turns it into behaviour. Adoption is what you see when the new way of working sticks, and it follows from the other three.
A team can begin building this internally, and expert-led training shortens the learning curve, builds the same standard across the team instead of in one careful person, and ties practice to the workflows you actually run. If you’d like to design these controls together on real work, that’s what our corporate workshops are for. For the manager’s side, see Managing the Responsible Use of AI.
A monitor who isn’t being measured tends to drift into glancing at the dashboard. Counting logins won’t help here, because oversight isn’t about whether people open the tool. Use three questions in order, and write the before figure down first.
| Question | What you look at | Example for an on-the-loop refund or reply workflow |
|---|---|---|
| Use: is the check happening? | The scheduled review, not the tool | Samples read on the days planned; interventions and pauses logged |
| Persistence: still happening a month later, unprompted? | Whether the monitor and their deputy run it without a reminder, and whether stop conditions are still current | The same person read the sample in week five; the thresholds were updated after the last rule change |
| Impact: is the work better or safer? | Sample catch rate (wrong items found per items read), time from first bad output to stop, errors reaching customers against the figure before | Catch rate trending down as fixes land; time to stop shortening; fewer reopened tickets than last quarter |
Write the before figure down first. A flat sample catch rate of zero means either a very good process or a sample that isn’t looking in the right place; test it with a few planted examples.
The catch rate deserves a second look. When it’s high, you’ve found a problem worth fixing. When it falls to zero and stays there, it’s tempting to relax the sample. The better move is to check that the sample is still looking in the right place, and to treat a rule, model or prompt change as a reason to go back to a tighter sample for a while. NIST’s AI Risk Management Framework, a voluntary framework, builds measurement into its Measure and Manage functions for the same reason.[6]
The record itself can be as plain as a log the monitor fills in after each review. Here are two rows, so you can see the columns and the standard.
| Date | Workflow | Items read | Wrong found | Pattern | Action | Decision |
|---|---|---|---|---|---|---|
| Week 1, Monday | Refunds under the cap | Sample as planned | One approved outside the policy window | Wording of the window is unclear in the policy document | Policy wording fixed; note sent to the policy owner | Stay on the loop; sample again Friday |
| Week 1, Thursday | Routine FAQ replies | Sample as planned | None | None | None | Stay on the loop; review thresholds after the next product change |
Illustration of the columns, not real data. Add a column for the person who read the sample, and keep the log where the deputy can see it.
That measurement record is also what lets a workflow move from in the loop to on the loop, and back. The rule that keeps it honest is below.
Move a workflow from in the loop to on the loop only after the sample record shows it’s earned it, and move it back when the model, the rules or the prompt change.
You don’t need a programme to start. You need one workflow and an hour.
Here is what the card looks like for the refunds workflow, filled in. The two rows people tend to leave blank are the pause rule and who restarts it.
| Field | Refunds under the cap |
|---|---|
| Control | On the loop under the cap; in the loop above it |
| Monitor and deputy | Customer support team lead; deputy is the finance manager |
| Review slot | Fifteen minutes each morning, in the calendar |
| Sample | About one in ten each week, or at least twenty, whichever is more |
| Alerts | Refund volume or average size jumps; a refund policy document is edited |
| Pause rule | Two wrong approvals of the same kind in one sample, or a change to the policy or the model |
| Who restarts it | The monitor, after a fresh sample of at least twenty reads clean |
A copy-ready card. The sampling numbers are starting values to adjust with your own record, not benchmarks.
This week, take the AI task your team trusts without checking and ask one question about it: who would notice if it went wrong, and how long would that take?
It’s when the person acts. In the loop, they approve each output before it takes effect. On the loop, the AI acts within limits while they monitor, sample and can stop it. A single workflow can use both, split by a threshold such as the size of a refund.
Not by default. Approving every item can fail when volume is high and reviewers start to defer. Monitoring can fail when errors are hard to see or can’t be undone. The better question is which failure your work can afford, and whether the monitoring has a sample, alerts and a stop condition behind it.
Decisions about a person’s job, pay or record, anything that can’t be taken back once it’s sent or paid, public statements you can’t retract, and anything a law, contract or policy says a person must decide. Check the rules for your own jurisdiction with counsel.
They read a sample on a schedule, look at the exceptions or alert queue, check anything over a threshold and pause the workflow when a stop condition is met. They also update the rules and thresholds after a change. It works best as a fixed slot in the week, not as something done when time allows.
Yes, in steps. Start with a person approving everything, keep a record of what they catch, and loosen one threshold at a time when the record supports it. Move back when the model, the prompt or the rules change. A workflow that was reviewed closely in month one hasn’t earned unattended running in month two.
I read the primary or official sources for this guide on November 2, 2026: the European Commission’s AI Act Service Desk text of Articles 14 and 26, the Department of Defense’s Directive 3000.09 as published by Washington Headquarters Services, the Congressional Research Service primer on US policy for lethal autonomous weapon systems, NIST’s AI Risk Management Framework page and the NIST Generative AI Profile (AI 600-1). The CRS primer is a secondary summary of the directive and is the source for the three loop terms. The directive text I read does not use them and is the 2012 version with the 2017 change, while the primer reports a January 2023 update I did not read in full. I read the first part of the NIST profile and not its final pages. Nothing here is legal advice, and which rules reach you depends on where you operate. The control-selection test, tables, sampling numbers and role guidance are Future Factors’ working tools, not standards. I haven’t tested any specific product, and the scenes are illustrations, not client stories.