Reliability Is The Product

The model is not the product. The agent is not the product. Reliability is the product.

Generative AI has become extraordinarily capable. New models arrive constantly, benchmarks improve, Context windows grow and agents become more autonomous. Yet none of that guarantees that an AI system will produce a dependable outcome when people actually rely upon it.

In Reliability Is the Product, Harrison Kirby presents a deeply practical, first-hand view of what production AI has taught him. Drawing on years of building and operating AI systems across government, policing, defence and industry, he argues that the hardest problems now sit beyond the model: defining what good looks like, evaluating against real requirements, controlling the evidence models receive, inspecting the seams between components, limiting complexity and agency, understanding the technology you depend upon, and proving that the complete system delivers real value.

This is not a neutral survey of AI engineering, nor a catalogue of fashionable patterns. It is an opinionated account of what works, what fails and what Harrison believes engineers should focus on when the system has to move beyond the demonstration and become something people genuinely depend upon.

Because a compelling demonstration is easy.

A dependable system is not.

Do not show me how intelligent the system is. Show me why I can depend on it.