The initial hysteria around artificial intelligence has settled. The era in which business leaders marvelled simply at a machine’s ability to produce human-like text is ending, and in its place is a more demanding one: implementation. Organisations have stopped asking “what can AI do?” and started asking “how does AI solve our exact, highly regulated, specialised problem?”

That is a much harder question, and it is where most pilots stall. A model that drafts a brilliant marketing email will fail badly when asked to analyse a complex contract or interpret anomalies in industrial sensor data. The gap is not a matter of prompting harder. It is architectural — and it separates generic, off-the-shelf AI from trained, domain-specific AI.

The generalist: a jack of all trades

Generic AI means the large foundation models — GPT-class systems, the standard Claude and Gemini assistants — built deliberately to be generalists. To achieve that range they are trained on internet-scale data: nineteenth-century literature, Python source code, cooking recipes and pop-culture trivia, all in the same weights.

Their great strength is immediacy. There is no setup, no training corpus to assemble, and they can pivot between unrelated tasks inside a single conversation. For brainstorming, drafting, summarising and general administrative support they are genuinely excellent and very cheap for what they do.

But that breadth is bought at the direct expense of depth. A general model predicts the most plausible next word given broad human discourse, which means it is optimised to sound coherent rather than to be correct. Pushed into a narrow technical field, that tendency produces hallucination: the model states a fact, a citation or a figure with complete confidence because nothing in its design requires the answer to be grounded in a verified source.

The specialist: a master of one

Domain-specific AI inverts the strategy. Rather than teaching a model everything about the world, you train and constrain it against a narrow, curated, often proprietary body of material: millions of financial statements, or a vision model that sees nothing but microscopic fractures in aerospace components.

A trained system trades general knowledge for accuracy in one field. You cannot ask a diagnostic model to write a haiku about the weather — but you can rely on it to flag an obscure anomaly, with a lower false-positive rate than a human specialist. Restricting the model’s universe to verified material is precisely what makes hallucination rare: there is far less room to invent when every answer must come from a known source.

There is a second benefit that tends to decide procurement. Because these systems do not need to carry the whole internet, they are smaller — small enough to run on infrastructure you control, which is how data residency and compliance obligations get met in practice rather than in a policy document.

So which one should you buy?

It is a question about stakes, not about technology. If the cost of a wrong answer is low — an off-tone first draft, a clumsy summary — generic AI is the obvious and economical choice, and paying for anything more specialised would be waste.

If a wrong answer means a missed diagnosis, a contract signed on bad terms or a decision taken on a fabricated figure, then the precision of a trained system stops being a luxury.

The organisations getting the most out of this will not choose one or the other. They will run hybrid architectures: a generic model as the conversational surface, routing anything consequential to an ecosystem of small, trained, domain-specific agents working quietly underneath. The generalist takes the question. The specialist answers it.