Most AI service failures are not model failures. They are organisational ones.

I've spent more than ten years working on customer experience from the technology and strategy side — the part where a service promise either survives contact with real systems, real content and real people, or quietly doesn't.

Most of that time has been spent in the gap between what a company says it will do for customers and what its infrastructure will actually let it do. I've worked on the platforms, the routing, the knowledge, the measurement, and the operating models that sit underneath support at scale. The interesting problems have almost never been the ones the org chart said they were.

I currently lead customer AI at Xero, a team I formed to work out how large language models change service delivery in practice rather than in demo. That work has been the clearest confirmation of something I'd half-believed for years: the constraint on great service was rarely effort or intent. It was cost, latency and the sheer impossibility of knowing every customer's context at the moment they needed you. Those constraints have moved. Very few companies have caught up to what that means.

I write here mostly to think out loud about that gap. Views are my own, examples are generalised, and nothing here is a statement on behalf of my employer.

I'm glad to talk with people building in customer experience and applied AI — comparing notes, arguing about measurement, or working through a problem you're stuck on.

Reach me at hello@bengilmore.online.