> In our most successful projects we are able to reach 95% accuracy

Is this the 95% of issues that are common between customers that could be and are handled at scale with a FAQ/knowledgebase that is surfaced to a user automatically through prompts?

How does it actually do on the smaller percentage of questions that are complex?

And that 95% is on the most successful projects. But what does the distribution of accuracy look like across all projects? If 2 projects out of 1000 are 95% accuracy that doesn't mean much to me.