01From decisions to systems

Beyond the Demo argued that generative AI makes the technology affordable and the management harder. It set out what must be decided before a GenAI system reaches its users, and by whom. Those decisions are necessary. They are not sufficient.

A decision that exists only in a governance document is not enforced. "Every answer cites the right version of its source" remains a sentence in a policy until something in the system records versions, carries them through to the answer, and checks them. "Answers never bypass access rights" remains a sentence until rights are taken from the source systems, kept current, and applied every time an answer is assembled. Governance that holds over years has to be built, tested and operated like any other part of the product.

This is also why so many pilots are rebuilt before they go live. A pilot is usually built to show that the idea works. It runs on a sample of documents, with no access rights, no record of which version produced an answer, and no way to tell whether a change made things better or worse. Each shortcut is reasonable for a demonstration. Together, they mean the pilot cannot become a production system. It can only be replaced by one.

This article describes what to build so that a pilot can grow into production instead. It uses the same example as the first article: a rail maintenance company, a composite, whose technicians need reliable answers drawn from decades of manuals, service bulletins and notes about locomotives that stay in service for thirty to forty years.

Two principles run through everything that follows. The first is to start from the problem and the data, never from the tool. A common failure is to pick a library or framework believed to be the best, then make everything else fit around it. The right choices depend on your documents, your users and your constraints, and the only way to know which choices are right is to measure them.

The second is that the business decides what is true and what is acceptable, and technology makes those decisions hold. What a term means, which source is validated, how much risk a question carries, who may see what: these are decided by the people who own the knowledge and its use. The engineering work is to record those decisions, enforce them, measure them and keep them over years. It is demanding work, but it serves decisions it does not make.