How to Keep a Human in the Loop With an AI Agent

Human in the loop means an AI agent does the work, but a person decides at the moments that matter: what the agent may touch, which actions need a yes, when a conversation needs a teammate, and whether the results were right. The goal isn't to watch every step. It's to put people at the few points where a mistake would be costly.

Get it wrong one way and the agent makes changes nobody would have approved. Get it wrong the other way and your team approves fifty routine actions a day and stops reading them. This guide shows where a person belongs, how to sort actions by risk, and how to keep approvals meaningful instead of a rubber stamp.

The strongest control is the one that makes an action impossible. If an agent should never cancel subscriptions, don't give it the ability and rely on instructions; leave the capability off. Then use instructions for judgment calls.

• Capabilities: Tick only what the job needs. An agent without a capability can't use it, however it's asked.
• Channels: Each agent answers only the channels you assign.
• Instructions: Write "check with a person first" and "hand off when" lists in plain words, like you would for a new hire.
• Roles: Customer agents work with customers; operator agents work with your team. Keep them separate.

Not every action deserves the same check. Sort them by what happens if the agent gets it wrong:

In Concierge, asking is the yes. When a customer texts to cancel their subscription, it's cancelled. When a teammate asks an operator agent in admin chat to cancel an order, it's cancelled. The agent asks first only about what it thought of itself, worded as an offer: a customer asks to cancel their latest order, the agent sees it came from a subscription, and it asks whether to cancel the subscription too. A scheduled task is different, because nobody is present when it runs: its instruction doesn't count as a request, so a step that could do damage needs a teammate's standing permission or an approval, and the question goes to the teammate who owns the task, in the task's thread. Nobody has to find an approvals screen; they answer in their own words, like "yes, go ahead" or "only the shipping."

Some conversations need a person, not a yes. A good handoff is fast, reaches someone even outside business hours, and arrives with context so the customer doesn't repeat themselves.

• Clear triggers: The customer asks for a person; the case is outside policy; the customer is upset about a repeat problem; anything involving damaged goods, missing packages or legal threats.
• A useful note: The agent checks what it can first, then hands off with what it found.
• Honest waiting: The agent tells the customer help was requested, without promising someone is online right now.
• One owner: When a teammate accepts or replies, the agent steps back. When they're done, they hand it back.

On a phone call, a teammate can listen in and join from the Finn apps. Read more in AI escalation and human handoff and the handoff docs.

• Read the record: Every request, approval, tool use and reply is saved in the conversation. If an answer surprises you, ask the agent why; it explains what it looked at and did.
• Review a sample every week: Read conversations from each channel start to finish, plus every handoff.
• Approve what it learns: After your team handles a handoff, Concierge agents suggest what they learned, such as the answer your team gave. You can require a person to approve each suggestion before agents use it.
• Check emails that wait: When an email can't go out without a person's OK, it shows as an Email Approval card in the conversation, with Approve & send or Reject.

An approval nobody reads is worse than none, because it looks like oversight. A few habits keep them honest:

• Don't ask about reading: Lookups should never need a yes. Save attention for changes.
• Make the ask specific: A good approval says what will happen, to which order, and what the alternatives are.
• Remove instead of approving forever: If you always say no to something, turn the capability off.
• Put the right person at the point of decision: An offer goes to the person who asked; refunds go to the team who owns the money.
• Revisit monthly: As the agent proves itself on a job, give it more of that job; when something goes wrong, narrow it.

Concierge builds the human into the work. You choose what each agent can do with approvals and capabilities, what someone asks for runs, what the agent thinks of itself is offered first, refunds stay with your team, handoffs reach your team in the team inbox and the Finn app, and every step is on the record. Concierge Starter is $20 a month with 20,000 credits. Pro is $199 a month, or $1,990 a year, with 199,000 credits. More credits cost $1 per 1,000, phone and text charges are billed as used, and there are no per-seat fees. See pricing.

Glossary: human in the loop, approval workflow, audit trail, AI governance, least-privilege access and human handoff. Read next: how to roll out an AI support agent in four stages, the AI agent security checklist, AI agent vs chatbot and what is an AI employee?

Human in the loop means the AI agent does the work, and a person decides at the points that matter: what it may do, which actions need a yes, when to take over, and whether the result was right.

Anything that moves money, and any change the agent decides on by itself. Reading data, answering questions and doing what a person explicitly asked for shouldn't need a separate approval.

No. A Concierge agent makes the changes someone explicitly asks for without asking again, and asks first only about what it proposes on its own. Refunds stay with your team. If you never want an agent to take an action, turn that capability off and it can't.

No. Once the customer proves who they are with a sign-in code, the change they asked for runs without another question. Refunds are the exception: the agent opens the return and your team approves the refund.