Home / Glossary / Human in the loop
Glossary
What is human in the loop?
Human in the loop is a control where a person approves an AI action before it happens. The AI stops at a set point, shows what it wants to do and waits for a yes or a no. In AIBound's AgentHarness it is the "ask" in allow, ask or deny.
By the AIBound team · Updated 17 November 2026
Key takeaways
- Human in the loop means a person approves an AI action before it happens.
- It fits actions that are risky but sometimes needed, such as reading a file that holds credentials.
- An agent that asks about every action teaches people to click yes without reading, so the question belongs on the few actions that matter.
Why human in the loop matters for AI agents
An AI agent can run commands, read files and send data with the employee's access. Most of what it does is harmless. A few actions are hard to undo, such as deleting a folder or sending a file to an outside host. Human in the loop puts a person's judgment in front of those actions.
It also gives security a middle choice. Denying every risky action stops real work, and allowing every action leaves no check at all. Asking keeps the work moving and still puts a person in front of the decision.
How human in the loop works
The company chooses the actions that need a person. The agent does the rest on its own.
- 1 The security team decides which actions need approval, such as destructive commands or access to secrets.
- 2 The agent reaches one of those actions and pauses before it runs.
- 3 The person sees what the agent wants to do and answers yes or no.
- 4 The agent carries on or stops, and the answer is kept as a record.
Human in the loop vs human on the loop
Both keep a person involved. The difference is timing. In the loop, the person decides before the action. On the loop, the AI acts first and the person watches and can step in.
| Question | Human in the loop | Human on the loop |
|---|---|---|
| When does the person act? | Before the action runs | While or after the action runs |
| What does the AI do while it waits? | It pauses until the person answers | It keeps working unless the person stops it |
| What is it best for? | Actions that are hard to undo | Routine work that a person reviews later |
Sources
- OWASP LLM06: Excessive Agency · genai.owasp.org
- OWASP LLM01: Prompt Injection · genai.owasp.org
Sources checked 19 November 2026
Examples
Human in the loop examples
FAQ
Frequently asked questions
What is human in the loop?
Human in the loop is a control where a person approves an AI action before it happens. The AI pauses at a set point, shows what it wants to do and waits for an answer. It is used for actions that carry risk, such as deleting files or reading secrets.
What is the difference between human in the loop and human on the loop?
In the loop, a person approves the action before it runs, and the AI waits. On the loop, the AI acts on its own while a person watches and can step in. In the loop is slower and safer. On the loop suits routine work that someone can review later.
When should an AI agent ask a person for approval?
Ask when an action is risky but sometimes needed: a destructive command, access to a credential file, or data sent to an outside host. Low-risk actions can be allowed and clearly harmful ones denied. Too many questions lead to approval fatigue, where people click yes without reading the request.
How does AIBound put a human in the loop?
AgentHarness, an AIBound product, enforces policy inside the agent with three outcomes: allow, ask or deny. With ask, the action is put to the user as a question before it runs. The security team writes the rule in plain English and applies it to the company, a team or one laptop.