AI-Native Architecture

Cloud-native produced a shared substrate because every application team was rebuilding the same primitives badly. The same consolidation is starting for AI: the gateway, the evaluation harness, the retrieval plumbing, the prompt registry, the routing controls. Which of those belong to a platform team, which belong to the feature teams, and the rule that separates them.

Cloud-native forced DevOps into existence because the old split between building and running stopped working. AI-native opens three ownership gaps: evaluation, the context supply chain, and the routing and cost decisions. A new title is usually how an organization avoids a change rather than makes one. Here is who should own each gap, and the one role that might actually be new.

The cloud era had the six R's, and they mattered because they gave teams permission to leave most of the estate alone. AI-native needs the same thing more, because the pressure to declare everything AI-native comes from outside engineering. Four criteria that sort the estate, and three answers instead of one.

The distributed monolith was the cloud-native failure that looked like success. AI already has its own catalog: the load-bearing prompt, the eval that tests the model instead of the system, RAG on a data model never meant to be retrieved from, the agent with tool access, and the special-case integration. Each one, and the check that finds it.

Five levels, from "we added a chat feature" to "this is how we build now," each with an observable test instead of an adjective. Most organizations are two levels below where they believe they are, for the same reason they were on cloud maturity. The demo works, so the architecture must be fine.

AI components belong on the same risk matrix as everything else. What changes is how you score them: a likelihood rubric built on how often a person meets the failure, and a severity rubric built from reach, reversibility, and detection lag.

Services version on your schedule. Models drift, get deprecated, and improve on someone else's. Designing the boundary around a component you do not control, and the honest math on how much abstraction is worth paying for.

Retrieval, freshness, and data boundaries stop being back-office plumbing the moment a model depends on them. The context supply chain deserves the same rigor we learned to give service contracts, and it is where AI security actually lives.

Distributed systems changed architecture because the network made a function call expensive. AI does it again, except the cost is money as well as latency. Model routing is an architectural decision, not an optimization you do later.

Every testing, monitoring, and reliability practice you own rests on one assumption: same input, same output. An LLM in the call path removes it. What replaces it is evaluation, thresholds, and verification as an explicit architectural layer with a cost and an owner.

Every enterprise says it's adopting AI, the same way every enterprise once said it was in the cloud. Bolting AI features onto an existing system is the new lift-and-shift — the four-properties test for what genuinely AI-native architecture requires instead.