Depending on the use case, a 10% failure rate can be quite acceptable. This is of course for non-critical applications, like e.g. top-of-funnel sales automation. In practice, for simple uses like labeling data at scale, I'm actually reaching 95-99% accuracy in my startup.