This deep dive expands on the Loop Asia conversation with Karin Verspoor, Executive Dean of Computer Science at RMIT University.

Every industry has its own version of what Karin Verspoor spent her early career untangling in medicine. Financial services splits into retail, institutional, corporate, and wealth. Telecom keeps network, billing, and customer-experience domains that don't share a data model. None of it is a healthcare problem. It's what happens in any domain complex enough to need experts, handed to a computer that doesn't have any.

The forty-year head start

The Unified Medical Language System is decades of work by the US National Library of Medicine, starting from a cataloguing problem, not an AI one: the International Classification of Disease organizing conditions hierarchically, SNOMED capturing clinician language for signs and symptoms, UMLS tying the two together by shared meaning. By the time electronic records and generative AI arrived, healthcare had forty years of it to build on.

Even dictionary lookup — mapping terms straight to UMLS vocabulary — worked as a starting point, because the vocabulary investment was already made. Not sophisticated. Available. That's what most enterprises chasing an AI mandate don't have. There is no UMLS for your business — if it exists at all, it's scattered across data dictionaries, tribal knowledge, and whatever the last integration project reconciled before the budget ran out.

Negation is the tell

The failure mode Verspoor described — a system reading "no evidence of infection" and flagging "infection" as a positive finding, because it matches keywords rather than meaning — is a diagnostic test, not just a healthcare anecdote. Ask what your own systems do with negation, and the same failure usually turns up in a support ticket classifier, a contract review tool, a customer sentiment model. It's not a model problem, it's a context problem: you can't get negation right without a structured understanding of what a term relates to.

Large language models hold that context far better than a lookup table ever did. But they haven't trained on the full diversity of any specific domain's documentation — there's a gap between what a model knows in general and what your business needs. Closing it is domain-specific work, through structured vocabulary and relationship-mapping, not around it.

The mandate without the roadmap

This is the shape of the problem a lot of technical leaders face: a board directive to deploy AI, and no one owning what context it needs or where that context lives. It's tempting to treat that as a model-selection problem, or a prompting problem. Verspoor's career says otherwise — the hard part was never getting a model to generate text, it was building a structured representation of a domain's vocabulary solid enough that any tool could reason over it correctly. That's an architecture decision, not a purchasing decision. It's where a technical champion needs a peer who's done this governance and integration work — not a vendor selling the next layer of tooling.