One of the watershed events that marks a successful R&D project is the handover to engineering.
This step generally involves a fair amount of integration effort, and occasionally requires that the code be rewritten to employ databases, web services or conform to data contracts with other modules. Sometimes, this also requires that the logic be rewritten in a different programming languages, perhaps a translation from Python to C# or from Perl to Java.
I’ve noticed in most such projects, even after the most faithful reimplementation of the code, the output only matches to three decimal places. Beyond that, the results invariably vary (pun intended).
For most decision support application (which dominate the artificial intelligence field), this is close enough for all intents and purposes. However, for applications from the field of finance, certain life-sciences and space technologies, this would be considered a minor crisis!