MCP and AI basics
Human in the loop
A design where a person reviews or approves an AI action before it happens. The default safeguard for spend changes, refunds and customer messages.
Human in the loop means a person reviews or approves an AI action before it takes effect. In MCP setups it is usually implemented by the host, which can pause before a tool runs and show what is about to happen, and by the way you design a workflow, for example asking the assistant to draft a change rather than make it.
It is the default safeguard for anything with a cost of being wrong: budget and bid changes, price updates, refunds, customer messages, deleting records. These are also the actions prompt injection would aim at, so the approval step protects against manipulation as well as against ordinary mistakes.
Approval only works if the person can see enough to judge. Ask for the proposed change and its reasoning together, keep write tools to a few named tasks, and avoid blanket 'always allow' settings for servers that can change live data. For reporting and audits, no approval is needed, which is why read-only connections are where most teams start.