An intelligent system is any system whose behavior is shaped by models — retrieval pipelines, agents, recommenders, autonomous decision engines. When such systems sit in the request path of customer-facing products, they inherit the operational obligations of infrastructure: capacity planning, change management, rollback, and SLOs.
The teams that ship intelligent systems successfully treat them as software. Prompts are versioned. Evaluations are CI checks. Model swaps go through canary deployments. The exotic becomes routine.
Design for reversibility
The single most valuable property of a production AI system is reversibility — the ability to undo a decision, roll back a model, or contain a misbehaving agent without taking the product offline. Reversibility is a design constraint, not a recovery procedure, and it changes how systems are architected from day one.
Adaptive without unaccountable
Adaptation and accountability are not opposites. Systems can learn from production traffic while keeping a complete audit trail of what changed, when, and on which evidence. The discipline is in the instrumentation — and in resisting the temptation to ship adaptive behavior that no one can later explain.