The development of product-service innovation projects within the context of a company is not yet supported by clear theories and methodologies. Our objective is to analyze innovation and idea generation for such projects from the fuzzy front end to the selected design concept, assessing their potential to be successfully developed and launched on the market. We present a protocol study, using which data derived from 19 innovation projects of five types and conducted by 86 students are analyzed. Sixty-one variables are observed, thus generating 700 data vectors. Bayesian network learning is used to explore conditional inferences among these variables. We examine conditional probabilities between the innovation process means and the significant results produced for the company, modulated by the influence of contextual variables. A number of surprising findings are drawn about the link between problem setting and problem solving processes, the importance of certain contextual variables, and the potential discrepancies between the apparent and produced results of innovative projects. Conducted analyses imply the need for novel innovation evaluation frameworks.