Most enterprise software today has AI sitting on top of it, not inside it. Pull out the chatbot, and the product still works the same way. That is the bolt-on failure mode, and it describes the majority of AI deployments since 2022.
The core argument here, built around AWS as a concrete case study, is that products carry two learning layers: conceptual overhead invented by the builders to organize their own systems, and interaction overhead to navigate the interface. Bolt-on AI reduces friction in both layers but does not eliminate the conceptual debt. A user asking the AWS assistant about storage still learns what S3, buckets, and IAM are. The AI teaches the product's language instead of replacing the need for it. The distinction between domain knowledge users must own versus product-invented abstractions users should never have to touch is the analytical core of this piece, and it is worth reading for that framework alone.
The alternative model proposed is AI as a permanent translator between user intent and product primitives, absorbing concepts like accounts, services, and resources so users never surface them, not just at setup but through errors, capacity issues, and ongoing management. The piece is careful to separate what AI should abstract from what it cannot: domain knowledge stays with the user, product overhead does not. The design implication is structural, not cosmetic. If the products being built now still require users to learn builder-invented vocabulary, the AI layer is decorative.
[READ ORIGINAL →]