Discussions of AI today tend to put it all in a single basket. Large language models, machine learning, and agentic systems are interchangeably used, although they address entirely different problems. I have witnessed teams spending a ton of money on the wrong layer and then wondering why their systems are not moving anything to production.
The majority of AI systems do not crash immediately when a team realizes that something has gone amiss. The outputs are reintroduced into the model, the dashboards remain green and the operations will still be running until a decision goes awry, the accuracy plummets or anomalies begin to emerge in workflows. It is hardly the model that is the issue. The issue in the realism is that the pipeline systems that exist cannot match the operational complexity in real-time.
With the continued AI revolution, it is becoming more and more obvious that conventional approaches to the management of Reliability, Availability, and Governance (RAG) do not apply to the modern AI architecture. The RAG 1.0, as we have known it, was useful in the earlier days of the AI integration, however in a rapidly changing world of decision making, it is desperately in need of something much more strong.