Human in the loop AI agents: where people approve, and how
By the Codivated team · Updated
Human in the loop means a person reviews or approves an AI agent's work before it takes effect, at the points where a mistake would cost money, trust or compliance. In practice that is a short list of actions, an approval screen people will actually use, and a log. Most agents start with a person approving everything and earn autonomy action by action.
What does "human in the loop" mean for an AI agent?
It means the agent prepares the work and a person makes the call at defined points. The agent drafts the reply; a person sends it. The agent prepares the refund; a person approves anything over $200. It is different from "human on the loop", where the agent acts and a person monitors and can step in, and from full autonomy, where nobody checks individual actions.
| Mode | The agent | The person | Use it for |
|---|---|---|---|
| Human in the loop | Prepares each action | Approves before it happens | Money, customer-facing messages, record changes, early weeks of any agent |
| Human on the loop | Acts on its own | Watches, samples and can take over | Proven, low-risk actions at volume |
| Autonomous | Acts on its own | Reviews reports | Reading, sorting, internal drafts |
Which actions should need a person?
Sort every action the agent can take into three tiers: on its own, ask a person, never. The deciding questions are simple: Can it be undone? Does it move money or change what a customer sees? Is there a rule a person would apply that the agent might not? Our governance guide walks through the tiers in detail.
What approval patterns work in practice?
Our team has built all four of these in production, across customer-facing agents that have taken more than 2.5 million autonomous actions:
| Pattern | How it works | Best for |
|---|---|---|
| Approve in Slack or Teams | The agent posts the proposed action with context and Approve / Deny buttons | Internal agents, small teams, fast decisions |
| Review queue in a dashboard | A queue of pending actions behind your company login, with filters and bulk approve | High volume, back-office work like orders and invoices |
| Customer approval | Your customers see the agent's proposed work in your product and approve, deny or ask for changes | Agents that act on a customer's behalf, like posting content |
| Escalation and takeover | The agent works on its own until a rule or its own uncertainty hands the case to a person, who can take over and the agent stops | Conversations, like lead follow-up and support |
For example, our team built lead-nurture agents that worked conversations on their own and escalated to a person when a lead was qualified or the agent wasn't sure, and a person could take over a conversation at any time. We also built content agents where customers approved, denied or requested changes on every post before it went out.
How do you keep approvals from becoming a bottleneck?
- Show the reason, not just the action. "Refund $84: damaged on arrival, photo attached, within the 30-day policy" gets approved in seconds.
- Batch the routine ones. Let a person approve 20 clean matches at once.
- Widen by rule, not by feel. When a category has been approved unchanged for a few hundred cases, let it through on its own below a limit.
- Track approval time. If approvals wait hours, the agent saves nothing.
What doesn't work?
Putting every rule into the agent's instructions and trusting it to follow them. Past a few dozen rules, models stop following all of them reliably. The rules that matter, like limits, permissions and never-do actions, belong in code that checks the agent's output before anything happens. Focused checking steps can review quality. The approval screen is the last line, not the only one.
Is human oversight required by law?
For some uses, yes. The EU AI Act requires human oversight for high-risk AI systems, and frameworks like NIST's AI Risk Management Framework expect clear roles for human review. Even where it isn't required, an approval step and a log make an agent far easier to trust, audit and improve.
Common questions
Doesn't human approval cancel out the time savings?
No. Approving a well-prepared action takes seconds; doing the work took minutes or hours. And approval is temporary for most actions.
Who should approve an agent's actions?
The person who would have done or signed off on the work before, at the same level of authority.
How does the agent know when it isn't sure?
Through checks you can measure: missing fields, values outside normal ranges, conflicts between sources and low confidence on a match. Those route to a person automatically.
Sources
- Article 14: Human oversight, EU Artificial Intelligence Act.
- AI Risk Management Framework, NIST.
- Building effective agents, Anthropic: checkpoints for human feedback in agent design.