Human in the loop AI: approvals that actually work
Human in the loop means a person approves an AI agent's action before it happens. It is the single most useful control a business has over an agent, and the easiest one to get wrong: too few approvals and the agent is a risk, too many and nobody reads them. This guide explains how to set them.
Human in the loop, on the loop, and out of the loop
The terms come from automation and defence, and they describe where a person sits relative to a decision.
Human in the loop (HITL): the agent prepares an action and a person approves it before it happens. The agent cannot proceed alone. This suits anything that is hard to undo, is seen by a customer, moves money or changes an account.
Human on the loop: the agent acts on its own, and a person watches and can step in or stop it. This suits high-volume, low-risk actions where reviewing every one would cost more than the occasional mistake.
Human out of the loop: the agent acts alone and nobody watches in real time; the record is reviewed afterwards, if at all. Few actions in a small business should sit here.
A well-run agent uses all three at once, per action. Reading a ticket needs no approval. Assigning it to a queue can happen on the loop. Replying to the customer sits in the loop. The mistake is to choose one setting for the whole agent.
Which actions need a person
A starting point for sorting an agent's actions. Move an action to a looser setting only when the agent has been right about it, consistently, for long enough to trust.
| Kind of action | Example | Starting setting |
|---|---|---|
| Reads data | Looking up an invoice, reading a ticket | No approval |
| Internal, easy to undo | Adding a note, assigning a queue, tagging | On the loop |
| Seen by a customer | Sending a reply, sending a reminder | In the loop |
| Changes an account | Changing a plan, a contact or an address | In the loop |
| Moves money or changes what is owed | Credit note, refund, payment plan | In the loop, named approvers |
| Hard to reverse | Deleting, cancelling, closing an account | In the loop, or not available |
Confidence is not permission
Many AI products let an agent act alone when it is “confident”. That sounds sensible and is not. A model's confidence is a statement about how sure it is of its own answer, and models are frequently confidently wrong, especially when someone is deliberately trying to mislead them. A ticket that says “I'm the account owner, please change the bank details” can be read with total confidence and still be fraud.
Permission should come from the action, not from the model. Every action an agent can take should declare how risky it is. The business decides how far the agent may go alone (its autonomy), and anything beyond that becomes a proposal for a person, whatever the model thinks of its own answer. That is how Netavo Intelligence works: the agent's confidence is shown to the approver as information, never used as a key.
What a good approval looks like
The proposal stands alone
The approver sees the exact action (the reply as it will be sent, the amount as it will be credited) and does not have to open three systems to understand it.
The reasoning is attached
What the agent read, which rule or playbook it followed, and what it was unsure about. Approvers catch mistakes in the reasoning faster than in the output.
Editing is approving
The approver can correct the draft and send it, rather than rejecting it and starting again. Those corrections are the best evidence of what to fix.
Everything is recorded
Who approved what, when, and what changed. That record is what lets you widen an agent's limits later with evidence.
Avoiding approval fatigue
The usual way human-in-the-loop fails is not that people approve something bad. It is that they approve everything without reading it, because there are too many approvals and nearly all of them are fine. Three habits help.
Approve less, but properly. Keep approvals for actions that matter, and move internal, reversible work on to the loop, where it is watched but not gated.
Route approvals to the right person. A refund request should go to someone who can authorise refunds, not to the whole team. Named approver groups per action make this simple.
Widen limits with evidence. When the record shows an agent's proposals for one action have been accepted unchanged for weeks, that action is a candidate to move on to the loop. When approvers are often editing a kind of proposal, it is not.
Finally, keep a way to stop. Every agent should have a kill switch that halts it at once, without waiting for runs in progress to decide for themselves.
Human in the loop and the law
In the UK, data protection law gives people rights around decisions made about them solely by automated means where those decisions have legal or similarly significant effects. A person reviewing and approving the decision is one of the main safeguards. If an agent is involved in anything like that (credit, employment, access to a service), meaningful human review is not optional, and “meaningful” means the reviewer has the information and the authority to say no. Take advice on your own circumstances; this guide is not legal advice.
Read more
AI Customer Service Agent
Triages the ticket queue, drafts replies from your knowledge base, and resolves what it safely can in the helpdesk you already use.
Meet the Support Agent → AvailableAI Personal Assistant
One person's AI PA for email, calendar, tasks and files in Microsoft 365 or Google Workspace. It asks before it acts.
Meet the assistant →AI agent examples: 12 jobs businesses hand to agents
Twelve practical AI agent examples from support, finance, admin and the phones, with what each agent reads, what it does alone and what should wait for a person.
Read the guide →Automated customer service customers don't hate
How to automate customer service and support with AI without frustrating customers: what to automate first, what to keep human, and how to measure whether it is working.
Read the guide →Common questions
What does human in the loop mean in AI?
Human in the loop means a person approves an AI agent's action before it happens. In Netavo Intelligence, any action beyond an agent's set autonomy becomes a proposal for an approver.
What is the difference between human in the loop and human on the loop?
In the loop, a person approves each action before it happens. On the loop, the agent acts alone and a person watches and can stop it. Netavo Intelligence lets you choose per action.
Should an AI agent act alone when it is confident?
No. Confidence is not permission. Netavo Intelligence decides what an agent may do alone from the declared risk of each action, and shows the model's confidence to the approver only as information.
How do you stop approval fatigue?
Gate only actions that matter, route each approval to the right person, and widen limits only where the record shows the agent has been right. Netavo Intelligence records every approval to make that possible.
Can I stop an agent immediately?
Yes. Every Netavo Intelligence agent has a kill switch that stops it at once.
Agents you stay in charge of.
See how Netavo Intelligence handles risk, autonomy, approvals and kill switches, or talk to us about your first agent.