Approvals: staying in control of what AI does

What this is

The question every owner asks first is "what stops it from doing something stupid?" In Korps the answer is an approval: a held action that does not run until a person says yes. The worker asks; you decide; the record shows both.

Who it helps

Anyone letting AI touch real systems: send email, update a CRM, pay an invoice, post to a channel. The more consequential the tool, the more this page matters.

How an approval works

1. A worker reaches an action that is marked as needing approval, for example sending an email through your connected Gmail.
2. Korps holds the action. It does not run.
3. A Needs You item appears: what the worker wants to do, to whom, and why. It waits across devices and restarts, so you can answer from your phone an hour later.
4. You approve or decline. On approve the action runs once, exactly as shown. On decline the worker records it and continues without it.
5. Activity records the request, the decision, who made it and when.

Everything that is not yours goes elsewhere. A blocked worker goes to its team lead. A stalled or failed task goes to its manager. Only a decision that genuinely needs a human reaches you.

Where to draw the line

Three places you can set it, from broad to narrow:

• By kind of action. Require approval for whole categories: read, write, send, delete, execute. A common starting point is "send, delete and execute always ask; read never does". This organisation-wide policy is an Enterprise edition control today; on other editions the same effect comes from the next two.
• By tool. When you connect a business tool, each worker's access to it can be marked "requires approval". Your CRM steward can update records freely; your outbound worker must ask before every send.
• In the assignment. Tell the worker the specific cases: "Ask before any refund over the usual amount."

Start strict. Loosen one action at a time once you have seen the worker handle it well for a few weeks.

A concrete example

A sales team is qualifying inbound leads. The Researcher can read the CRM and the web freely. The Outbound worker drafts every reply but its connection to email is marked "requires approval", so each send arrives as a Needs You item with the draft attached. You read it on your phone, approve, and the email goes. After a month you notice you have approved every follow-up without edits, and you relax the rule for follow-ups only, keeping it for first contact.

Requirements and availability

• Available in Korps. Approvals, Needs You, per-worker tool grants that require approval, and the Activity record are built in.
• Organisation-wide action policies (approve every "send", deny a tool for everyone) are an Enterprise edition feature; ask before relying on them.
• Approvals apply to actions taken through Korps connections. See limitations.

Limitations and approval points

• An approval governs actions the worker takes through Korps: connected tools, browser sessions, runtime tools it reports. A worker's own reasoning is not an "action" and cannot be approved step by step; the assignment and the model choice govern that.
• Approving means you have read what will happen. A Needs You item you approve without reading is a decision you made.
• Declining does not undo earlier steps. Set approval before the consequential action, not after it.
• Korps does not hold your model provider credentials or your tool passwords. Connections are yours, and you can remove one at any time.