Michael Finlan

Abstract Building a deterministic delivery model around a probabilistic component is the central architectural challenge of using LLMs in production. In this session, I present a governed pattern where the LLM is constrained to produce structured, schema-bound change proposals inside a host framework. Those proposals are validated, reviewed, and only applied after explicit approval. The LLM becomes a contributor, not an executor. I will cover why curated admissible context can outperform agentic discovery in high-control environments, why markdown is a weak control surface for precision work, and how validation gates and typed outputs and artefacts change the reliability profile of LLM-assisted delivery. This session is for architects, data engineers, and technical leads who want a rigorous, implementable way to use LLMs in production without treating them as autonomous actors. *This talk would work well as a two-session deep dive — one covering the methodology, one showing development using it. In that case the second session would be titled "Governed AI Delivery in Practice".