SCNET · Enterprise IT · Ankara, Türkiye

Sanal Çekirdek

A bot that keeps customers waiting is worse than a queue that does.

An assistant is judged by how many customers reach a resolution, not by how many questions it answers. Sanal Çekirdek builds around the handoff threshold: where the assistant stops and a person takes over.

An assistant that answers wrongly costs more than one that does not answer at all. Saying it does not know is better behavior than guessing.

Scope and handoff threshold

An assistant does not try to answer everything. Which topics it covers and when it hands over are written at the start; handing over is part of the design rather than a failure.

  • Out-of-scope topics are recognized and handed over at once
  • A customer repeating themselves twice triggers handoff
  • Conversation history transfers with the handoff
  • A customer can ask for a person at any point

Accuracy and sources

An answer given to a customer counts as a commitment by the company. Answers therefore rest on approved content and can cite a source; where no source exists, the assistant produces no answer.

  • Answers come from an approved content set
  • Pricing and commitment topics are handled separately
  • No source found means handing over instead
  • Wrong answers return to the content owner

Language, tone and accessibility

The assistant speaks in the company's voice. A tone guide is written; abbreviations are expanded and legal text is not pasted verbatim. Accessibility belongs to the design too.

  • The tone guide is written with sample answers
  • Equal capability is targeted in both languages
  • Screen reader usability is tested
  • Long answers are broken into readable pieces

Measurement and feedback

Success is measured by resolution rate and by the experience after a handoff. Conversations that start with the assistant and land with an agent are tracked separately — they are the richest source of improvement.

  • Resolution rate is reported by topic
  • Waiting time after handoff is measured separately
  • Satisfaction is asked separately for assistant and agent
  • Unanswered topics enter the content plan

How we work

  1. Write the covered topics and the handoff threshold
  2. Prepare the approved content set
  3. Build the tone guide with examples
  4. Launch in narrow scope
  5. Track resolution rate and widen scope

How success is measured

  • Resolution rate is known per topic and rising
  • Waiting time after handoff stays short
  • No answer is produced where no source exists
  • Wrong answers turn into content corrections

Frequently asked questions

Will it replace our agents?

That is not the goal. Repetitive and simple questions stay with the assistant while agents move to more complex work. The handoff threshold exists to protect that split.

Who is responsible if it gives wrong information?

The company is. That is why answers come from approved content, pricing and commitment topics are excluded, and every answer ties to a source.

How is this different from the Generative AI page?

The technical side belongs there: model selection, guardrails and cost. This one covers the customer-facing side: scope, handoff, tone and resolution rate.

Share your ten most frequent customer questions; we will see together which of them the assistant can keep.

Share your ten most frequent questions