Decision card
A decision card is the human-in-the-loop moment in AI support. When the agent is not cleared to act alone, it neither sends a blind answer nor gives up. It presents a human with a compact card containing everything needed to decide in seconds.
What is on the card
- The customer request and the relevant account context.
- The agent's recommended reply or action.
- Its reasoning and the sources it retrieved.
- Its confidence and which guardrail tripped.
- Clear choices: approve, edit, send, escalate, or reject.
Why it trains the model
Every decision a human makes on a card is a training signal. Edits become preference pairs that teach the preferred phrasing, approvals confirm a pattern is safe to automate, and escalations become new cases. The queue gets smaller over time because the AI learns the judgment it watched a human apply. Escalation is a rewarded outcome, not a failure.
Frequently asked
When does Resolve show a decision card instead of auto-resolving?
Whenever confidence does not clear the bar or a guardrail trips: a low-confidence retrieval, an angry ticket, a forbidden action like a large refund or a chargeback, or an enterprise account held to a higher bar. The card gives a human the context to decide fast.
Does using decision cards slow support down?
The opposite. The card front-loads the request, context, recommendation, and sources so a human decides in seconds, and each decision reduces how often that pattern needs a human next time.