What is human-in-the-loop AI?
Human-in-the-loop AI is a design in which an AI system produces work but a person makes the decision that carries consequence. The AI drafts, classifies, summarises or proposes; a human reviews that output and approves, edits or rejects it before it takes effect. It is not the same as a notification or an activity feed, which tell a person what already happened. In a genuine human-in-the-loop system the process stops and waits: the message is not sent, the quote is not issued and the purchase is not placed until an approval is recorded. Well-built implementations add three properties. The pause is durable, so a pending approval survives a restart or a deploy. The audit trail attributes every change to a human or to an agent. And a human edit takes precedence, locking the record so that a later AI run cannot silently overwrite the correction.
Approval gate vs notification
| Notification | Approval gate | |
|---|---|---|
| When it fires | After the action | Before the action |
| If nobody reads it | The action still happened | Nothing happens |
| Undo | Damage control after the fact | Not required |
| What it protects | Awareness | The customer, the price, the order |
The distinction matters commercially. A quote sent at the wrong margin, or a cold email sent to the wrong list, cannot be recalled. Awareness after the fact is not a control.
What a correct implementation requires
A durable pause
An approval may sit for days. If the pending state lives only in memory, a restart either loses the work or re-runs earlier steps and duplicates their side effects. The pause has to be persisted so the run resumes exactly where it stopped.
Actor-attributed audit
Every change records whether a human or an agent made it, and which agent run produced it. Without that, no one can reconstruct why a record looks the way it does, and "the AI did it" becomes unfalsifiable.
Edit precedence
When a person corrects an AI-produced field, that field is locked against further AI writes. Otherwise the next run quietly reverts the human's judgement and the reviewer stops trusting the queue.
A named approver role
Approval is a permission, not a courtesy. The role that can release a quote should be explicit, and separate from the role that can merely view or draft.
Where WorkAgents AI applies it
All three products share the same rule: AI drafts, a person approves.
- AI Outreach — nothing sends without approval; drafted emails and replies wait in an approval queue.
- Quote AI — a drafted quote stops at a review gate before it can be issued to a customer.
- Agentic ERP — agent proposals across CRM, procurement and inventory wait for an approver, and every mutation is audit-logged.
Related
Decide where your approval gates belong
We map which steps in your process can be automated and which must stop for a person.
Book a Free Workflow Audit