We are looking to automate our customer support tickets using a generative ai workflow. The goal is to draft responses automatically for human review. Does anyone have experience with the ethical guardrails needed here? I'm worried about "hallucinations" appearing in official company emails and damaging our brand reputation.
3 answers
Integrating a generative ai workflow requires a "Human-in-the-Loop" (HITL) approach as your primary safety net. You should never let the AI send an email directly without a person clicking "approve." Additionally, use RAG (Retrieval-Augmented Generation) to ground your model in your company's actual documentation. This ensures the ai workflow pulls facts from your knowledge base rather than making things up. We implemented this in late 2023, and it reduced our manual drafting time by 60% while keeping our "hallucination" rate near zero.
Are you planning to fine-tune a model on your historical tickets, or are you going with a zero-shot approach? The complexity of your ai workflow will vary significantly based on that decision.
Focus on strict prompt engineering first. A well-structured prompt in your ai workflow can often outperform a poorly fine-tuned model and is much cheaper to maintain.
Michael is spot on. We found that adding "Check your work against the provided manual" to the ai workflow prompt solved most of our accuracy issues immediately.
We are leaning towards fine-tuning on about 10,000 resolved tickets. Bradley, do you think that's enough data to keep the ai workflow consistent with our brand voice, or should we rely more on prompt engineering and few-shot examples instead?