Guides9 min read

Human in the Loop AI: What It Means and How to Approve Agent Actions From Your Phone

Human in the loop AI explained: where people fit in an AI system, how it differs from human on the loop, and how to approve agent actions from your phone.

By The Unorderly teamPublished

Quick answer

Human in the loop AI means a person takes part in how an AI system works instead of leaving it entirely to the model. For AI agents, that usually comes down to one rule: the agent can research and draft on its own, but costly, public or irreversible actions wait for a person's yes. A phone notification is the quickest place to give it.

That's the short version. The longer one covers where the idea comes from, how "in the loop" differs from "on the loop", and a practical setup for approving agent actions without being stuck at your desk.

What is human in the loop AI?

The term is older than AI agents. IBM defines human-in-the-loop (HITL) as "a system or process in which a human actively participates in the operation, supervision or decision-making of an automated system." Stanford HAI puts it more simply: humans "may provide guidance, correct errors, or make final decisions to improve the accuracy and reliability of AI."

People can sit in the loop at three different points, and most articles blur them together:

  1. Training. Humans label data so a model can learn, for example tagging images as "car" or "bus". Google Cloud's explainer also describes active learning, where the model picks which examples it wants a human to label.
  2. Feedback. Humans rate or rank outputs, and the model is tuned on those ratings. Reinforcement learning from human feedback (RLHF) is the best-known version.
  3. Decisions at run time. The system pauses before acting and a person approves, edits or rejects the action. This is the one that matters if you run agents.

If you use Claude, Grok, Cursor or OpenClaw to do real work, the first two happened long before you showed up. The third is yours to design.

Human in the loop vs human on the loop vs out of the loop

Related searches for "human in the loop" are full of "human on the loop", and the difference is practical, not academic.

Mode What the person does Does the system wait? Good for
Human in the loop Approves or rejects specific actions before they happen Yes, until there's a decision Sending as you, spending money, deleting, publishing
Human on the loop Watches what the system is doing and steps in if something looks wrong No, it keeps going Routine, reversible work with a visible log
Human out of the loop Nothing while it runs; maybe reviews results later No Reading, research, summaries, internal drafts

You don't pick one mode for the whole agent. You pick one per action. A sales agent can be out of the loop for looking up leads, on the loop for updating your pipeline, and in the loop for emailing a customer.

Why human in the loop matters more now that AI agents act

A chatbot that gets something wrong gives you a bad answer. An agent that gets something wrong sends the email. That's the shift.

The OWASP Top 10 for LLM Applications calls this risk "Excessive Agency" and recommends that you "utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken." It adds that this can live "in a downstream system (outside the scope of the LLM application)", which is the part most people skip. Hold on to it, because it comes back in the mistakes section.

Regulation points the same way. Article 14 of the EU AI Act says high-risk AI systems must be designed so they "can be effectively overseen by natural persons." It specifically warns about "automation bias", meaning people trusting the system's output too readily. Your inbox agent probably isn't a high-risk system under the Act. The lesson still applies: oversight only counts if the person can actually judge what they're approving.

The agent tools already work this way. In Claude Code's Manual permission mode, Claude "stops and asks you before most actions that edit files, run shell commands, or reach the network." Developer frameworks such as the OpenAI Agents SDK and LangChain have built-in ways to pause a run until a person decides.

The catch is that most of those prompts appear in a terminal or chat window on the computer where the agent runs. When you walk away, the agent waits. One common complaint in agent communities is approval queues stalling all day while the owner is away from their desk.

Where the loop should sit for an AI agent

You want to be in the loop for a short list of actions, not all of them. Gate everything and you'll start tapping yes without reading, which defeats the purpose. A useful default:

  • In the loop: anything that sends a message as you, spends money, deletes or overwrites data, publishes something, or changes who has access.
  • On the loop: reversible changes inside your own tools, like moving a task or tagging a lead, as long as each change is labelled with who made it.
  • Out of the loop: reading, searching, drafting, summarising.

For the full set of tiers and a two-question test for borderline actions, see our guide to human-in-the-loop AI agent approval.

Step-by-step: how to approve agent actions from your phone

This works whether you build agents in code or use Claude, Grok, Cursor or OpenClaw off the shelf.

  1. List the actions your agent can take. Go through every connection it has: email, calendar, CRM, ad accounts, files, social accounts. Remove access it doesn't need before you design any approvals.
  2. Mark the in-the-loop actions. Use the list above. Be strict about anything that sends, spends, deletes, publishes or grants access.
  3. Split proposing from doing. The agent prepares the action (the email body, the recipient list, the refund amount) and asks. The step that actually carries it out runs only after a yes. If the agent can do the action whether or not you approved it, your approval is just a suggestion.
  4. Send requests somewhere your phone shows them. That might be a notification from an approval app, a chat channel you check, or a review queue on a dashboard. Pick one place so requests don't get lost across three apps.
  5. Give every request the same shape. Tell the agent in its instructions what each request must contain: the exact action, who it affects, the cost, the reason, and whether it can be undone. You should be able to decide from the lock screen without opening a laptop.
  6. Decide what yes, no and silence mean. Yes lets the action run. No comes with a short note so the agent can try again. Silence is never a yes: give requests an expiry and let stale ones lapse.
  7. Check the record weekly. Look at what you approved and rejected. If you approve one type of action every time for a month, consider moving it to on the loop. If you keep rejecting one, fix the agent's instructions.

What a good phone approval request looks like

A phone screen is small, so the request has to do the work. Compare:

Weak request Strong request
Action "Email the churned customers" "Send this email (text below) to these 3 addresses"
Reach and cost Not stated 3 recipients, no cost
Reason Not stated "They cancelled this week; this is the win-back template you approved last month"
Undo Not stated "Can't be unsent"
Expiry None "Skip if not approved by Friday"

If you can't decide without opening another app, the request is missing something. Tell the agent what to include next time.

Mistakes to avoid

  • Rubber-stamping. Threads like "human in the loop is becoming corporate theater" exist for a reason. Too many prompts and people stop reading. Gate fewer, bigger things.
  • Approving the summary, not the action. "Email the churned customers" can turn into thirty recipients. Approve the actual recipient list and text.
  • Letting the agent check its own approval. If the same agent that asks also decides whether you said yes, a mistake or a prompt injection can skip the check. OWASP's advice to enforce approvals "in a downstream system" is about exactly this.
  • Mistaking the record for the gate. An approval app records your decision. Something still has to read that decision before the action runs, whether that's your agent's instructions, a workflow tool or a small script. Make sure that piece exists.
  • No history. If you can't answer "who approved this, and when?", you can't learn from mistakes or explain them to a client.

How Unorderly helps

Unorderly is a command center for iPhone, iPad and Mac where AI agents such as Grok, Claude, Cursor and OpenClaw publish dashboards, to-dos, pipelines and approval requests. You connect your agent once with a link (our remote MCP server explainer covers what that means, and the custom connector guide shows the steps for Claude). After that, the agent can put approval requests next to the numbers it already tracks.

When it needs a yes, you get a notification you can answer from the lock screen, on any plan. Unorderly records your approve or reject decision, along with your note if you said no, and your agent reads it and carries on. You can also have the decision sent to your own webhook, which the agent can't see. Unorderly doesn't run or block the action itself: your agent or your webhook does that, which is why step 3 above matters. Every view shows which agent wrote it and when.

Unorderly is free to download on the App Store. Free covers 1 agent, and Pro is $9.99 a month or $79.99 a year. If you're still choosing tools, the AI agent dashboard buyer's guide explains the difference between apps that mirror an agent's session and apps that show its results.

FAQ

What is human in the loop AI?

Human in the loop AI is any AI system where a person takes part in the process instead of leaving every decision to the model. That person might label training data, rate outputs, or approve an action before it runs. For AI agents, it usually means the agent proposes and a human says yes or no before anything high-stakes happens.

What is the difference between human in the loop and human on the loop?

In the loop means the process waits for a person: nothing goes ahead until someone approves it. On the loop means the system runs by itself while a person watches and can step in. Out of the loop means no person is involved while it runs. Most sensible agent setups mix them: on the loop for routine work, in the loop for anything costly, public or hard to undo.

Is human in the loop the same as agentic AI?

No. Agentic AI describes systems that plan and take actions on their own. Human in the loop is a design choice you can add to those systems, so that certain actions wait for a person. The two work well together: the agent handles the work and you decide on the short list of actions that matter.

Can I approve AI agent actions from my phone?

Yes. The agent has to post its request somewhere your phone can reach, such as a notification, a chat message or an approval app, and your answer has to get back to whatever runs the action. The phone is the easy part. What matters is that the request shows the exact action and that nothing runs until your decision is recorded.

Does human in the loop slow AI down?

It adds waiting time to the steps you gate, which is why you should only gate a few. Reading, research and drafting can stay automatic. If approvals reach your phone, the wait is usually however long it takes you to look at a notification, not until you get back to your desk.

Does the EU AI Act require human in the loop?

Article 14 of the EU AI Act requires high-risk AI systems to be designed so that people can effectively oversee them, including being able to override or stop them. It doesn't use the phrase "human in the loop" or require a person to approve every output. Whether it applies to you depends on whether your system counts as high-risk, so check the Act itself or get legal advice.

About Unorderly

Unorderly is an app for iPhone, iPad and Mac where AI agents such as Grok, Claude, Cursor and OpenClaw publish live dashboards: KPIs, pipelines, to-dos and approvals, built from fourteen native widgets. Agents connect through one remote MCP server with a single sign-in, and approvals can be answered from the lock screen. It is free to download on the App Store; every new account starts with 14 days of Pro, and Pro costs $9.99 a month or $79.99 a year. Unorderly is made by Swiftful Dynamics in Australia.