Most AI programmes stall on ownership, not technology. A central team builds a platform nobody asked for, or every department buys its own tool and the organization ends up with ten incompatible ones. An operating model settles who decides what.
A workable split
- The centre owns the shared layer — approved models, security patterns, evaluation tooling, vendor contracts and the register of live systems.
- Business units own the use case — the problem, the success measure, the people whose work changes and the budget for running it.
- Both own quality — the centre sets the method for testing, the business supplies the examples and signs off the results.
Rules that keep it working
- Anything touching customer data or external actions is registered centrally, whoever builds it.
- A business team can use the shared layer without a long approval, so long as it follows the published patterns.
- Every live system has one named owner on the business side, and one named technical contact.
If teams route around the centre to get things done, the centre is too slow. If the centre can't see what teams run, it is too weak. Fix whichever side is failing.