AI and agents
Human in the loop
Human in the loop is a step where a person approves, edits or rejects what an automation or AI agent proposes before it acts, used wherever a mistake would be costly.
How it works in practice
The point is to put judgement where it is needed and nowhere else. If every action needs approval, people stop reading and click yes. If none does, an agent can send the wrong price to a customer. We place approvals on actions that are expensive, irreversible or customer-facing, and let routine, reversible steps run.
A good approval step shows the person what the agent saw, what it proposes and why, in the tool they already use, such as Slack, email or a review queue. Their decisions are logged, and the share of proposals approved unchanged is a useful measure of when a step can safely run without review.
A worked example
An accounts team lets an agent match supplier invoices to purchase orders. Matches within tolerance post automatically. Anything over $5,000 or with a price difference goes to the finance lead in Slack, with the invoice, the PO and the difference shown.
Examples in this glossary describe typical builds for companies of 1 to 1,000 people. They are not named client results.















