Decision
The agent returns eligible, refuse, or review — not a free-text guess
Placeholder. A measured result has not been approved for publication.
Case study · AI engineering
A support agent prepares a refund decision and waits when it is unsure
Placeholder case study of an AI support triage flow. No client metrics are published.
Tech stack
- Ticket trigger
- Order and policy retrieval
- Decision policy
- Human review
This is an anonymised example. The client name and any measured result stay unpublished until the client approves them, so nothing below is a performance claim.
01 — The challenge
What did the client need?
Refund requests arrived as tickets. Agents re-opened the order, the policy, and the history in three places before deciding, and the decision was not recorded with the evidence.
02 — The solution
What did we build?
The ticket triggers an agent that retrieves the order, the policy, and the history, then returns a structured decision. Eligible cases can proceed to the refund action. Unsure cases stop for a person, with the context attached.
03 — Results we can describe
Results we can describe
Action
Only an eligible decision is allowed to start the refund path
Placeholder. A measured result has not been approved for publication.
Review
Unsure cases are handed to a person with the retrieved context
Placeholder. A measured result has not been approved for publication.
04 — Architecture
How the pieces connect
- 01Ticket opened
- 02Retrieve context
- 03Reason
- 04Act or review
Tech stack
- Ticket trigger
- Order and policy retrieval
- Decision policy
- Human review