Stop Building a Second Lakehouse Inside Your AI Platform: The Pattern, the Trap, and the Rules

Most enterprise AI platforms ship with a second lakehouse underneath the agents: ingest, bronze, silver, gold, governance, all copied from the company's own systems. Does the AI platform need to own that layer? This is an architecture decision record, narrated. Four options, ten criteria, every score tagged with where it came from (vendor documentation, practitioner reports, history, or my own judgement), five ways read-in-place breaks and the rule for each, a vendor-neutral reference architecture with ownership on the left edge, the preconditions, and a gated decision with revisit triggers. What you will take away: • Two tests that separate a modular AI product from a bundled one: open out, open in. • Why "zero-copy" means two different things, and which one is safe to build on. • Where the cost, residency and governance surprises hide, and how to govern them. • Five rules that keep indexes, caches and agent memory from becoming a data layer. • When read-in-place does not work, and what to do instead. Status, stated plainly: analysis of public documentation and practitioner reports. No client deployment. Scenarios are illustrative. No vendor is named or targeted. Chapters 0:00 The slice at the bottom 1:42 Two words, one phrase 3:16 The decision 4:59 What each side gets 6:32 Where it breaks 9:22 Derived state, not a data layer 10:21 Reference architecture 11:45 Preconditions and the record 13:04 Close #EnterpriseAI #DataArchitecture #Lakehouse #ZeroCopy #ApacheIceberg #DataGovernance #AIAgents #SolutionArchitecture