Human-in-the-loop AI: how to design approvals operators actually use
Human-in-the-loop only works if the approval step is fast, informed and honest. How to decide what an AI agent may do alone, what it must ask about, and how to keep approvals from becoming rubber stamps.
George Jucan, Founder, Poligon Tech 4 min
Short answer
Human-in-the-loop AI means a person approves, edits or rejects what an AI agent proposes before it takes effect. To make it work, sort actions by risk and reversibility, let the agent do low-risk reversible work alone, require approval for anything that commits money, changes a customer-facing record or can't be undone, and design the approval screen so a person can decide in seconds with the evidence in front of them.
“A person will approve everything” is the most common answer to the question “what if the AI is wrong?” It is also, too often, where the design stops.
An approval step is a piece of software like any other. Designed badly, it is either so slow that operators route around it, or so easy that it becomes a click with no thought behind it. Designed well, it is the reason an AI agent can be trusted with real work.
Here is how we design them.
Start with risk and reversibility, not confidence
The tempting rule is “auto-approve when the model is confident”. It doesn’t hold up. Models can be confidently wrong, and a confidence score says nothing about what a mistake would cost.
Sort actions on two questions instead:
- What does a mistake cost? Nothing, an annoyance, money, a customer relationship?
- Can it be undone? Easily, with effort, or not at all?
| Easy to undo | Hard or impossible to undo | |
|---|---|---|
| Low cost | Agent acts alone | Agent acts, person reviews afterwards |
| High cost | Person approves, can be quick | Person approves, with evidence and a reason |
In practice, that means:
- The agent acts alone on preparation work: gathering context, checking stock, matching documents, drafting.
- A person approves anything that commits money, changes a price or a credit term, sends something to a customer or supplier, or changes a record other people rely on.
- Nothing irreversible is automatic in the first months, whatever the confidence.
Confidence still has a job. It decides how a case is presented (obvious ones grouped for quick approval, uncertain ones flagged for a closer look), not whether a person sees it.
Design the approval screen for a decision in seconds
The approver is usually an operator in the middle of a busy day. The screen should let them decide quickly and correctly:
- Lead with the proposed action, in plain words: “Enter order for 40 × SKU 1182 at €12.40, deliver Thursday.”
- Show the evidence next to it: the line from the customer’s email, the contract price, the stock check. The approver should never have to open three systems to check the agent’s work.
- Highlight what is unusual: “Price differs from last order by 8%”, “First order from this delivery address”.
- Offer three actions: approve, edit, reject. Editing matters most. It turns “almost right” into “right” without starting over.
- Ask for a reason on reject or edit. One tap from a short list, plus an optional note. Those reasons are how the agent improves.
Illustrative example. Picture a wholesaler whose customer service team keys orders from emails and PDFs into the ERP. The agent reads each incoming order, matches products to SKUs, checks the price against the customer’s contract and checks stock. Clean orders appear in a queue with the original email beside them, ready to approve in one click. Orders with a price mismatch or a missing product code appear separately, with the mismatch highlighted and a suggested fix. The team still decides every order. They just stop retyping them.
Keep approvals honest
The real risk with human-in-the-loop is not that people reject too much. It is that, after weeks of good suggestions, they stop looking. Some ways to keep the check real:
- Track approval behaviour. Time to decide, edit rate and reject rate per approver and per case type. A queue that is approved in under a second with no edits for weeks needs attention.
- Seed known cases. Occasionally include a past case with a known problem and see whether it is caught.
- Rotate and limit. Long sessions of approving near-identical items lower attention. Batch by type and keep sessions short.
- Make “I’m not sure” an option. Escalating to a colleague should be one click, not a failure.
Log everything, for people and for auditors
Every proposal should be stored with:
- what the agent saw (the inputs and retrieved facts),
- what it proposed and why,
- who approved, edited or rejected it, when, and with what reason,
- what was finally written to the system.
That trace answers the auditor’s question, the manager’s question (“why did this go out at that price?”) and the engineer’s question (“where did the agent go wrong?”). It also becomes the evaluation set you use to test every future change.
Loosen the loop with evidence, not optimism
Approval rules are not permanent. When a category of case has been approved without edits for long enough, on enough volume, you can move it a step: from “approve before” to “review after”, say. Make that decision on the logged data, one category at a time, with the process owner’s sign-off, and keep the option to tighten it again.
What we don’t do is plan for full autonomy from the start. The goal is not to remove people from the process. It is to remove the retyping, searching and checking that keep them from the part of the work that needs judgement.
A note on regulation
The EU AI Act’s Article 14 requires effective human oversight for systems classed as high-risk. Many internal operations agents fall outside that category, but the same design principles apply, and building them in from the start is far easier than adding them later. For your specific case, ask legal counsel.
If you are designing an agent and want a second opinion on where the approvals belong, that is exactly what our workflow audit maps out: what the agent does alone, what a person approves, and what stays entirely human. For the technical side of the same boundaries, see What is an agent harness?
Frequently asked questions
What does human-in-the-loop mean in AI?
It means a person is part of the decision process: the AI prepares or recommends, and a human approves, edits or rejects before the action takes effect. It is different from human-on-the-loop, where the AI acts and a person monitors and can intervene afterwards.
Which AI actions should always need human approval?
Anything that commits money, changes prices or credit terms, sends something to a customer or supplier for the first time, deletes or overwrites records, or cannot easily be undone. Low-risk, reversible preparation work, such as gathering context, drafting or checking, usually doesn't need approval.
Does the EU AI Act require human oversight?
Article 14 of the EU AI Act requires human oversight for high-risk AI systems, as defined in the Act. Many internal operations agents are not high-risk, but designing for clear human oversight is still good practice and makes later compliance easier. Check your specific use case with legal counsel.
How do you stop approvals becoming a rubber stamp?
Show the evidence, not just the answer; require a reason when someone overrides; track how often each approver edits or rejects; and periodically insert known test cases. If approvals happen in under a second with no edits for weeks, the step is no longer a real check.