Refunds above a threshold
A support agent can answer, apologise and log all day. The moment it wants to move money, it asks the person whose budget it is.
Refund $840 to order 1183? Customer says it arrived broken; photos attached to the ticket.
Every one of these is the same call with different words in it. What changes is who you send it to and how long they have.
A support agent can answer, apologise and log all day. The moment it wants to move money, it asks the person whose budget it is.
Refund $840 to order 1183? Customer says it arrived broken; photos attached to the ticket.
Cold email, a reply to a customer, a post on the company account. The draft is the agent's; the send is a person's.
Send this 240-contact sequence? First touch drafted below.
Let the agent build, test and stage on its own. Gate the one command that reaches production, and record who waved it through.
Promote build 4f21c9 to production? 3 migrations, one of them drops a column.
Deleting a bucket, rotating a key, tearing down an environment. The kind of action where 'the model was confident' is not an acceptable postmortem line.
Delete staging-db-07 and its snapshots? Last write was 46 days ago.
An agent holding a card or a crypto wallet needs a ceiling it cannot reason its way past. Put a person above the ceiling.
Buy 12 months of the $4,800 plan? Cheaper monthly for the first 5 months.
On-call gets a Telegram message with two buttons and an hour to use them. No answer means no, and the runbook takes the safe branch.
Auto-scale the cluster 3× for the traffic spike? Est. $190 for the day.
Send one to yourself and watch the whole loop close.