AI on legacy systems: the work nobody demos

The model is usually the easy part of modern AI projects. The hard part is extracting useful data out of a twenty-year-old mainframe system without breaking the monthly financial close or operational integrity.

Across our last eight distinct projects, integration with legacy systems consumed between forty and sixty percent of total project effort. The variance depended on how many legacy systems had documented APIs.

An honest estimate in the first client meeting states that plainly and directly. It makes the proposal less attractive to clients. Clients hear forty percent of budget on infrastructure plumbing.

Projects that fail almost never fail on the model quality. They fail because data arrived late, or did not arrive, or arrived without the required permissions to access it.

Modern cloud-native architecture patterns are taught as universal defaults in education. Twelve years later there are still systems running COBOL and IBM mainframes in production.

Know the legacy footprint before estimating timeline and cost. Integration almost always takes longer than the shiny parts.