James draws a distinction that runs through the whole episode. Case context is the data of a single transaction: the customer, the claim, the payment in front of you. Policy context is everything that transaction should be judged against: regulation, policy, procedures and the tacit knowledge in experts’ heads. Most AI architectures flatten the two into one stream of tokens handed to a language model. That is where auditability dies.
Rainbird treats policy as computable knowledge, represented in a knowledge graph and reasoned over by a deterministic inference engine. The same case gets the same answer every time, with an evidence tree showing how the conclusion was reached.
“We’ve had 25 or 30 years of trying to get our data in order. Now we need to get our knowledge in order,” James says. Language models have a real part to play in that work: interviewing experts, synthesising documentation and drafting knowledge graphs that people then test and refine. What they should not do is make the decision. “Language models are exceptionally important, but they sit at the edges. LLM-as-judge is a bad idea.”
The conversation also covers the four failure modes of relying on language models alone: non-determinism, untestability, drift as policy changes over time, and the inability to produce a true account of how a decision was made. And it looks at who should own an organisation’s knowledge layer, with James making the case that business analysts, who have spent decades eliciting understanding from domain experts, are the natural knowledge engineers of the AI era.
There is also the story of where the Rainbird name comes from. Bertrand Russell and a certain wooden woodpecker are involved.
Listen to the episode
Listen to the podcast using your favourite streaming service.
