Choosing a representative sample
Include the awkward records: the oldest, the newest, the ones staff complain about, a month with a system change. A tidy sample produces a tidy prototype and an untidy surprise later. Strip personal data you do not need before sharing.
Why vendor demonstrations mislead
Demonstrations use the vendor's data, the vendor's scenario and the vendor's best operator. Ask for a trial with your data, your edge cases and your staff at the keyboard. Note how long setup takes; it is a preview of ownership cost.
Throwaway prototypes, and when not to throw them away
Decide before you start. A prototype built to learn should be discarded once it has taught you. If you suspect it will be kept, build the core with tests from the beginning and say so in the boundary.
Writing processing rules people can read
Name each rule as a sentence: “Orders without a customer reference go to exceptions.” Keep the code for each rule small and tested. Publish the list of rules with each output so readers know what was done.
Weighting criteria honestly
Ask stakeholders to weight criteria before seeing the options. Weights chosen afterwards tend to favour the option someone already liked.