Getting customer-facing chatbots right
Three things that matter most: a detailed system prompt, layered guardrails beyond the defaults, and fine-tuning on real, human-reviewed conversations.
Building customer-facing chatbots is always tricky because they’re open to all kinds of inputs. If not designed carefully, they can harm the brand instead of improving customer satisfaction.
Here are a few key points to get it right:
1. System Prompt – Make it detailed enough to reflect your brand’s voice, prevent prompt injection, and guide the conversation in line with your brand values.
2. Guardrails – Use default guardrails (e.g., filters for explicit content, hate speech, illegal activity) without limiting to them. Add custom guardrails and filters that fit your brand and address domain-specific risks.
3. Fine-Tuning – Instead of using base APIs alone, train your model on a decent volume of past conversations with human-reviewed annotations to create a consistent brand tone which represents the organisation.