The pattern is consistent across European industrial groups: a capable model, a convincing pilot, and no path into the systems that actually run the business.
Enterprise data does not arrive as a corpus. It arrives as transactions, master data with historical inconsistencies, plant-level exceptions and interfaces built over two decades. An assistant that cannot read those records with correct permissions, and cannot write back with an audit trail, remains a demonstration.
This is why the integration surface deserves the architectural attention normally reserved for models: identity and permission propagation, deterministic write paths, replayable event logs, and a clear boundary between suggestion and action.
The practical test for any enterprise AI proposal is simple. Ask what happens on the write path, who authorised it, and how it is reconstructed six months later during an audit.