Our summary
An AI assistant might propose refunding twelve euros when a customer asked for twelve cents. Markus Eisele shows how Java software can use Jev, a tool that answers set questions, to compare a proposed action with the customer’s original words. Permission checks and amount limits still belong in ordinary code.
LangChain4j connects those questions to Java methods. Jev receives the original request, the exact proposed action and relevant order details. Java then allows, rejects or holds that proposal for review. If its amount, order or action changes, it needs a fresh check; execution must also prevent duplicate payments.
Eisele compares two question designs on 24 English and 24 German cases. Separate checks for the order and amount wrongly approved a German request that explicitly forbade payment. His small exploratory test shows why matching details is not enough: the questions must also establish that the customer wants the action.
Key takeaways
- Compare the exact proposed action with the original request, not an assistant’s shortened account of it.
- Keep access rights, refund limits and payment execution in code; a high model probability is not permission to pay.
- Test incorrect approvals and cases held for review together; splitting a question can lose an important condition.