The Architecture Nobody Demos

The good news is that none of this means replacing what you already own. The data platform you invested in isn’t broken. It was built for a human reader, not for a machine asking questions in real time. The warehouse becomes the cold tier, the old chatbot’s intent taxonomy becomes the agent’s tool registry, and a new signal layer sits beside them.

The stakes rise most when users can transact from inside a conversation. People forgive a wrong answer, but not a wrong transfer. So confirmation has to be a structured screen, not a typed “yes,” and every action needs an audit trail linking it back to what caused it. That trail can’t be added later.

Written by Commencis Chief AI Officer Duru Kızılkayak, the third installment of the five-part series “The Signal Layer” covers the five early decisions that carry most of the architectural weight: what to do with your existing data platform, how streaming data enters the system, how memory is written and forgotten, how the agent’s output reaches the screen, and what changes when a conversation can execute transactions.