Superstitious, anecdotal reasons for not upgrading, OR known workload, trialled extensively on comparable proof-of-concept kit? Hide until 2017, OR clearly attempt to identify positives and negatives. Specific tests for specific features to get the best from the new version of the product, eg In-memory OLTP, Buffer-pool extension, cardinality estimator changes, writeable clustered columnstore etc in conjunction with your business-critical queries. Positive and negative behaviour changes identified. Workarounds identified for regressions where possible. Technical debt and/or connect items logged for items that cannot be resolved.

Evidence-based go/no-go decision based on the "balance". Maybe? : )