People often ask whether it's better to use a generalist model or one specialised in their field. The question assumes these are two competing products. In reality, they are two answers to different problems, and most organisations asking the question need neither — they need a third option.
We discussed this morning Thomson Reuters' legal model, and this trend will repeat itself in other sectors. Here's how to make sense of it.
What a specialised model actually brings
Three things, and they need to be distinguished because they don't all carry the same value.
Vocabulary. A technical field has its own terms, abbreviations and conventions. A model trained on it handles them more reliably. That's a real but modest gain, because the big generalists have already read a great deal of specialised literature.
Typical lines of reasoning. Every profession has characteristic sequences: how a lawyer qualifies a situation, how a doctor builds a differential diagnosis, how an accountant matches an entry. A specialised model reproduces these patterns better.
Grounding in a verifiable corpus. This is the most important gain — and, curiously, the one that doesn't require a specialised model at all.
The third option, which suits almost everyone
Here's the practical point. For the vast majority of organisations, what's missing isn't a model that knows the field — it's a model that consults the right documents.
A decent generalist hooked up to your document base via a RAG system produces answers grounded in your texts, quotable and verifiable. It doesn't know your trade; it reads your documents, which is often better.
This approach has three decisive advantages over a specialised model: it updates instantly when a document changes, it allows the source to be cited, and it respects each user's access rights. A model that has absorbed everything during training does none of these three things.
Three cases, and they're rarer than people think. Volume: if you're processing millions of identical tasks, a small specialised model costs far less per request than a large generalist. Offline use: if the system must run without a connection or in a fully isolated environment. Deep reasoning: if your work relies on sequences the model must master rather than consult, as with certain diagnostic reasoning.
The trap to avoid
A specialised model inspires a confidence it doesn't necessarily deserve. The fact that it's presented as an expert in a field encourages less checking.
Yet it remains subject to the same limits as all others, notably the absence of an uncertainty signal. A legal model asked about a jurisdiction it doesn't cover will answer with the same confidence. Its specialisation doesn't make it cautious — it makes it more credible, which is exactly the opposite of what you'd want.
What to take away
The useful question isn't generalist or specialised. It's: is what I need in my documents, or in the model's head?
If the answer is in your documents, what you need is a good retrieval system, not an expert model. If it requires deep professional reasoning you can't write down, then specialisation becomes relevant.
In both cases, the rule that doesn't change: the model proposes, a professional decides. That isn't a rhetorical precaution — it's the only arrangement that holds when no one can explain why a specific answer was produced.