A grip on answer quality
How do you make sure a resident always gets an answer that is not only correct, but also concise and genuinely usable?
Municipalities answer questions with Gem every day. Contact centre and web teams keep track of this and improve the logic and the answers where necessary, paying particular attention to the most frequently asked questions.
After 6 years, a great deal of experience has been built up. The database now holds almost 4,000 questions and answers. Recognised questions reliably receive the right answer. For questions that are not recognised, or are new, a member of staff is quickly on hand. That is how the service stays up to standard.
Using generative AI — AI sits an exam — Soundness comes first
There is also criticism (in Dutch) of chatbots, when questions come in that the chatbot has no good answer to. That is why we are working on using generative AI in Gem to improve this. Even so, that use remains a careful trade-off: we only apply it where it demonstrably improves the answers. We are testing several alternatives side by side, including GPT-NL, and comparing the differences in quality. We do this using practical ‘exams’, with questions selected in advance and matching model answers. And then we see which model scores highest!
Who will be allowed to give answers to residents?
Various language models are sitting an ‘exam’ at Gem to find out. Just like we once did at school!