Most AI programmes celebrate the first working prototype. The pilot ran, the output was useful and the technology is proven. What this establishes is that AI can work in this organisation's context. It does not establish that the organisation knows how to build AI workflows. Only that it has built one. An existence proof confirms that something is possible. It does not confirm that the methodology is repeatable, transferable or capable of scaling without the original team.
The second prototype tests a different hypothesis: is the methodology repeatable without the original team? If the same people, using the same undocumented process, build the second prototype in a different function, the result proves only that those people can do it again. The replication test requires the second prototype to be built by a different team, in a different function, following documented specifications from the first. If it succeeds, the organisation has demonstrated institutional methodology. If it requires the original team to intervene repeatedly, the methodology was never captured. It lived in the heads of the people who built the first prototype.
Organisations that move from individual capability to institutional methodology approach the second prototype with complete specification from the first: workflow selection criteria, governance requirements passed, configuration decisions and their rationale, human checkpoint design and handover procedure. This specification is given to the team building the second prototype, and deviations from it are tracked. Each deviation is a gap in the specification. The test is complete when the second team can build to specification without consulting the first team. That is the moment the organisation has demonstrated that its AI capability is institutional rather than individual.