AI strategies are written about models, use cases, and talent. They are delivered through the systems an enterprise already runs, and those systems set limits that no model can talk its way around.
The constraint beneath the roadmap
An agent that should update a maintenance record needs an interface to the maintenance system. An assistant that should answer from engineering documents needs those documents somewhere it can read them. When the system of record offers only a terminal screen or a nightly batch file, the AI program inherits that limit on day one.
This rarely appears in the business case. The cost shows up later, as integration work that takes longer than the AI work it was meant to support.
Three integrations usually matter most
Modernization does not have to mean replacing a core platform. For most AI use cases, a small number of integrations carry nearly all the value: the system that holds the transaction, the system that holds the reference data, and the system where people do the work. Giving those three clean, governed interfaces unlocks more than a multi-year replatforming promises.
Modernizing the right three integrations often matters more than choosing the right model.
Sequence modernization to the use cases
The practical approach is to let the AI portfolio set the order. Start from the use cases with the strongest value case, trace which systems they touch, and modernize those interfaces first. Each release then pays for a piece of modernization that the next release reuses.
AI can shorten the modernization itself
The same technology helps with the legacy work. Language models are useful for reading undocumented code, drafting tests around behavior nobody fully remembers, and mapping data between old and new schemas. Used with review by engineers who know the domain, this changes the economics of work that used to be deferred indefinitely.
Budget for it openly
- 01Name the legacy dependencies in every AI business case.
- 02Fund interface modernization as part of the AI program, not as a separate request.
- 03Build an API and integration layer that later use cases can share.
- 04Measure progress by use cases unblocked, not by systems retired.
An AI strategy that ignores the systems beneath it is a plan for pilots. One that budgets for them is a plan for production.