Human in the loop AI: how our agents ask before they act

Human in the loop AI means an AI system does the work, but a person approves the steps that matter before they happen. For a business, it is the difference between an agent you can trust with real customers and money, and one you are afraid to switch on. This guide explains how human in the loop works in practice, using the approval steps we built into our own AI agents: what each agent does on its own, what waits for a person, and how you decide where the line goes.

What is human in the loop AI?

In a human in the loop design, the AI prepares and proposes, and a person decides at fixed points. The agent still saves the time: it reads the request, gathers the facts, drafts the reply or prepares the change. The person spends seconds checking and approving instead of minutes doing the work.

Three patterns are common:

• Human in the loop: the agent waits for a person’s approval before a step runs.
• Human on the loop: the agent runs on its own while a person watches the results and can stop it.
• Fully automatic: the agent runs alone, used only for steps that are low risk and easy to undo.

A well-run agent uses all three, step by step. Sorting an inbox can run alone. Sending a customer email or changing a record waits for a person.

How our agents ask before they act: 3 answers

Every agent we run speaks to us in Microsoft Teams. When it reaches a step that needs a person, it posts an approval card: what it wants to do, the facts behind it, and the exact text or change it will make. The person chooses one of three answers:

• Approve: the agent carries out exactly what was shown, and logs who approved it.
• Modify: the person says what to change in plain words, and the agent redrafts and asks again.
• Reject: nothing happens, and the reason is recorded so the rules can be improved.

Approval is checked again at the moment the action runs. If something changed in the meantime, for example the customer already got a reply, the agent stops instead of acting on stale information.

What runs alone and what waits for a person

Here is where we drew the line in our own business.

Agent	Runs on its own	Waits for a person
Inbox triage	Reads each new email and sorts it: spam, newsletter, sales lead, license order, support or sponsored post	It never sends anything; spam is archived, never deleted
Sales Inbound	Creates leads, drafts the first reply, prepares license orders and test-runs them	Every reply to a customer and every order placed with a supplier
Finance	Categorizes and matches bank and card activity, drafts replies on billing questions	Anything that affects money; it never pays anyone
Marketing	Researches demand, writes and checks articles and social posts, prepares quotes from the shop’s prices	Every customer reply, every page or post before it is published, every social post
Client Services	Reads tickets and works out what kind of work each one is	What goes to the customer, approved by the person who owns the ticket

The rule underneath the table is simple: anything that affects a customer, money, a bid or a candidate is approved by a person. Everything is logged.

4 questions to decide where the human goes

When we design an agent with a client, we ask four questions about each step:

1. Who sees it? If a customer, a supplier or the public sees the result, a person approves it.
2. Can it be undone? A tag can be changed in a second. A sent email or a payment cannot.
3. What does a mistake cost? The higher the cost, the more the step needs a person.
4. Has it been proven? A new step starts with approval every time. Only after its results hold up do we consider letting a low-risk step run alone.

That last point matters. We start every new agent in shadow mode, where it proposes while people still do the work, and we only move a step to “runs alone” when the record shows it is safe.

The controls around the human

An approval step is only as good as the controls around it. Every agent we build also has:

• An audit log of every run, tool call and approval, so you can see who approved what and when.
• A pinned model and a daily cost cap, so behavior and spend stay predictable.
• Least-privilege access through a governed gateway, with no passwords held by the agent.
• An off switch for each agent, and one for all of them.

These are switched on, budgeted and audited from our AI Command Center, the layer that runs every agent we operate.

Common mistakes with human in the loop AI

Adding a person to the loop is easy. Making that person a useful check is harder. These are the mistakes we see most often:

• Approving without context. If the card only says “send reply?”, the approver cannot judge it. Show the exact text, the customer, the facts the agent used and what happens next.
• Too many approvals. If every trivial step needs a click, people start approving without reading. Keep approvals for steps that matter and let low-risk, reversible steps run alone.
• No way to say “almost”. A yes-or-no choice forces a bad draft through or throws good work away. A modify option lets the approver fix it in plain words.
• No check at the moment of action. If the situation changed after approval, the agent should stop. Re-check before acting, not only before asking.
• No record. An approval nobody can find later is no control at all. Log the card, the decision, the person and the time.

Get these right and human in the loop AI stays fast for your team and safe for your customers.

Why human in the loop AI is good governance

AI governance frameworks keep coming back to the same idea: meaningful human oversight of decisions that affect people. A human in the loop design gives you that oversight in a form you can show an auditor, a customer or your board: the approval card, the decision, the person and the time. The NIST AI Risk Management Framework and Microsoft’s guidance on responsible AI for Azure workloads both treat human oversight as a core control. We cover the wider picture in why responsible AI matters for businesses.

It is also how agents earn trust inside a company. Staff who approve an agent’s work every day see how often it is right, and they are the first to say when a step is ready to run on its own.

Start with one process, approved step by step

If you are weighing AI agents but worry about AI acting on its own, start where the risk is lowest: one process, every customer-facing step approved in Teams, everything logged. We explain how we build and run agents with people in control on our AI agents for business page, and our free AI readiness assessment will tell you which process is worth an agent first.

Frequently asked questions

Does human in the loop slow the agent down?

Only at the steps you choose. The agent does the reading, checking and drafting; the person spends a few seconds approving. For most processes that is still far faster than doing the work by hand.

Who approves the agent’s work?

The person who owns that process or record, for example the owner of a support ticket or the manager of a budget. Each agent has a named approver.

Can an approved step later run without a person?

Yes, if it is low risk, easy to undo and has proven itself. We move steps one at a time and keep the log, so the decision can be reviewed or reversed.

Is human in the loop AI the same as human oversight?

They are closely related. Human oversight is the goal: people stay accountable for what AI does. Human in the loop AI is one way to reach it, by putting an approval step in front of the actions that matter.

Where do approvals happen?

In our setup, in Microsoft Teams, where staff already work. The approval card shows the proposed action and the facts behind it, and the decision is recorded with the person and the time.

Does human in the loop AI work with Microsoft 365 Copilot?

Copilot works for the person using it, so that person is already in the loop for what they choose to send. For agents that act in back-office systems, the approval step has to be designed into the agent itself.

What happens if nobody approves?

Nothing is done. The card waits, and the agent never acts on a step that needs approval without one.