NatWest on agent safety: an 83% human-review pass rate hid a 54-point gap
AI Engineer · youtube · 2026-11-04
Jonathan Osborne, lead AI engineer at NatWest (UK's second-largest retail bank, 19M customers, 100K assistant conversations/day), explains how Cora is moving from scripted answers to AI agents: "building the agent is the easy part; proving it's safe isn't."
Key lessons:
- Blast radius is the real risk, not raw capability
- A routing funnel limits what ever reaches the model
- One accuracy score isn't enough—quality is several scores
- Human reviews: an 83% pass rate hid a 54-point gap
- An LLM judge scored far above humans with almost no correlation
- Shadow deployments with feature flags enable safe rollout, with human feedback as a gate
Highly relevant engineering practices for deploying agents in regulated industries.
More from coding & agent
- OpenAI DevDay demo shows forking agent task history at any step — paw_lean · 2026-11-04
- Veteran game dev: "AI builds a full game in a day" posts are mostly hype — bendee983 · 2026-11-04
- Dev builds interactive Excel-to-dashboard with ChatGPT, shares the prompt — TermRemarkable665 · 2026-11-04
- Agents cut dev time from weeks to a day; bottleneck moved to specs and review — Suspicious_Orchid770 · 2026-11-04
- Orgo, a YC-backed platform giving AI agents their own cloud desktops, claims 1,000+ teams — Scobleizer · 2026-11-04
- Founder: Vibe-coded a PMF-successful product, then hired pro devs to fix the mess — sujingshen · 2026-11-04