What does human in the loop AI actually mean?
Human in the loop AI means the machine does the work and a person approves the consequence. It drafts, a person sends. It scores, a person decides. It flags, a person acts. The loop is not a safety blanket bolted on at the end, it is a line drawn through the middle of the process at a specific step, and where you draw it determines both how much time you save and how much you can lose.
Most disappointing AI projects have the line in the wrong place. Either it is everywhere, so a person still touches every item and nothing is actually saved, or it is nowhere, so the first bad output goes straight to a customer.
Where should the gate sit?
At the last point before something becomes irreversible, and not one step earlier.
Irreversible means it left the building or it changed a record that matters. An email that reached a customer. A price that was quoted. A candidate that was rejected. A payment that went out. A record that overwrote the old value. Up to that point, a machine working alone costs you nothing but electricity if it gets something wrong, because you can throw the draft away.
Putting the gate earlier than that is the common error. If a person has to approve the machine reading the inbox, classifying the enquiry and drafting the reply, you have added three approvals to save one. Let it run the whole way to a ready reply, then have a person press send. One click, all of the work.
What deserves a gate and what does not?
Anything with legal, financial, safety or reputational consequence deserves a gate. Everything upstream of that usually does not.
- Always gated. Hiring and rejection decisions, credit and payment terms, pricing exceptions, anything touching safety, anything with a regulatory obligation attached, and first contact with a customer in a sensitive situation.
- Gated at first, then reviewed. Outbound replies, quotes, published content, data written back into your system of record.
- Rarely worth gating. Reading, sorting, classifying, summarising, drafting, flagging, enriching, and anything that produces a candidate answer nobody has yet acted on.
The test is simple. Ask what happens if this step is wrong and nobody notices for a week. If the answer is unpleasant, gate it. If the answer is that someone deletes a draft, do not.
How do you remove a gate safely?
By earning the right to remove it with evidence, over a defined period, on a narrow lane at a time.
Run the machine gated. Keep a record of every item where the approver changed something, and why. After a few hundred items you will have a number, not a feeling: the proportion of outputs that a person actually altered, and the pattern in the ones they did. If the machine is right on a well defined category and consistently wrong on another, you have just found your boundary. Let it run unattended on the first, keep the gate on the second.
That is a smaller decision than "should we trust AI", and it is a decision you can actually make. It is also why we take a baseline before anything changes, which is covered in how to run a 30-day AI pilot.
Who should hold the gate?
The person who would have been accountable for the output anyway, not whoever has time.
Approval is only meaningful if the approver has both the context to spot a bad answer and the standing to reject it. Handing sign off to someone junior because it is a routine click converts the gate into a rubber stamp, and a rubber stamp is worse than no gate at all because it manufactures a false record of oversight.
Give the approver two things: the machine's output, and the reason it produced that output. An approver who can see the working can catch a wrong answer. An approver looking at a confident paragraph with no provenance is guessing.
Does a gate defeat the purpose?
No, because the expensive part of most work is the drafting and the deciding, not the sending.
An owner who used to spend an evening a week writing quotes and now spends twenty minutes reading and approving them has not had their automation diluted by the gate. They have had the boring nine tenths removed and kept the judgement, which was the only part worth their time in the first place.
The gate also buys you something that pure automation never does: a person who still understands their own process. Businesses that automate a workflow into a black box lose the ability to explain it, and then cannot change it.
Design the loop before you build the machine
Decide where the gate sits, who holds it, and what evidence would justify moving it, and decide all three in writing before anything is built. That is a scoping conversation, not a technical one.
We put approval gates in by default and remove them only on evidence. If you want to see how that is written down before a build starts, our method covers it, and the crew pages show what each machine actually hands to a person. When you are ready to talk about your own process, get in touch.
Further reading worth your time: the Australian Government's AI Ethics Principles on human oversight and contestability, and business.gov.au for the general obligations that sit behind any customer facing process.