Whether to assemble it yourself or buy a product that does it. Buying is faster and constrains you to someone else’s roadmap and data terms. Building gives control and costs far more than the demonstration suggests, because the demonstration is about ten per cent of the work. The answer usually differs by layer rather than being one decision.
Technical view
Decompose the stack: model provider, orchestration and application logic, retrieval infrastructure, evaluation and observability, and interface. Buy where the capability is undifferentiated and mature; build where it encodes proprietary process or data advantage. Assess vendor products on data terms, model version stability, evaluation transparency and export capability. Note the middle option of building on an open framework, which trades vendor dependence for maintenance burden.
Related
Part of AI Terminology Index, section: Product and Adoption.