Building the AI Agent Is Easy. Proving It's Safe Isn't - Jonathan Osborne, NatWest

NatWest lead AI engineer Jonathan Osborne shares how the UK's second-largest retail bank, with 19 million customers and about 100,000 assistant conversations a day, is moving its digital assistant Cora from scripted answers to AI agents. Building the agent was the easy part; proving it was safe was the hard part. He covers the routing funnel that limits what ever reaches a model, why quality is several scores instead of one, human reviews where an 83% pass rate hid a 54-point gap, an LLM judge that scored far higher than humans with almost no correlation, and shadow deployments with feature flags for a safe rollout. Chapters: 0:00 Intro 0:41 NatWest and Cora 1:43 The real risk: blast radius 2:23 Scripted bot vs. AI agent 3:40 Four new ways agents fail 4:52 The routing funnel 6:34 Why one accuracy score isn't enough 7:43 Human reviews: 83% hid a 54-point gap 9:16 LLM judges that grade too kindly 11:06 How scoring runs in production 14:06 Shadow mode and gradual rollout 15:42 Human feedback as a gate 16:48 Lessons for regulated agents Jonathan Osborne https://www.natwest.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!