Need recommendation

A question for teams that have added AI to an existing SaaS or enterprise product. Did you keep the AI layer inside the main application, or separate orchestration, retrieval and model access into dedicated services? We usually prefer clear service boundaries when the system needs multiple models, strict permissions or independent scaling, but the added operational complexity is not always justified. Our team works on this type of enterprise AI integration. I would be interested in examples where a simpler architecture proved more reliable.