Yapay Zeka Destek Otomasyonu Testi

Testi çözdüğünüzde cevaplar ve açıklamalar görünür.

1. A support assistant answers a policy question. What should it show to help a customer assess the answer?
   • The current source used and a clear signal when information is missing
   • A confidence score without its source
   • A list of unrelated policies
   • Only the model’s version number
2. A customer asks to change an account setting. What should the system verify before acting?
   • Only the customer’s tone
   • That the model generated a fluent response
   • The correct account, requested change, authorization, and resulting state
   • That the same request appeared in a prior chat
3. Ticket closures rise, but reopenings also rise. Which interpretation is sound?
   • More closures prove support quality improved
   • Closure count alone may hide unresolved customer needs
   • Reopenings are unrelated to resolution
   • Automation rate is a complete success measure
4. What should a human handoff preserve?
   • Only the customer’s last word
   • The model’s private reasoning
   • A new unrelated ticket with no history
   • Relevant conversation and source context so the agent can continue the case
5. Why evaluate support across languages and uncommon problems?
   • Those cases cannot be supported by humans
   • Language never changes answer quality
   • Overall performance can hide failures for specific users or cases
   • A large average guarantees equitable service
6. A reply is generated from an old policy document. What safeguard is relevant?
   • Check source authority and freshness before sending the draft
   • Increase response length
   • Close the ticket automatically
   • Remove citations to avoid confusion
7. Why shouldn’t a model infer permission from a customer’s wording?
   • Models cannot understand requests
   • Authorization is needed only for refunds
   • Customer wording is always malicious
   • A request is not by itself proof of account authority
8. What can support teams learn from recorded corrections?
   • That the customer was necessarily wrong
   • Whether the answer or underlying knowledge base needs repair
   • That the model should never use sources
   • That all future cases can be automated