The success of a natural language dialogue relies heavily on the correct balance of several factors, such as fallback menus and visibility of the operator.  Large scale ‘Wizard of Oz’ testing on live traffic in call centres can help in determining the right balance.  This approach gives statistically valid data about customer behaviour and language in advance of costly engineering and natural language tuning cycles. Example data showing the lessons learned will be drawn from realworld use in scores of tests from Europe and the United States in recent years.