The debate on foundation models has made familiar an observation their very name suggests: they are a base, not a building (Bommasani et al., 2021). A generalist model has broad linguistic competence and shallow encyclopaedic knowledge; a system aimed at a professional domain needs something different, namely the operational knowledge of that domain: which documents circulate, which rules govern them, what counts as an error and who answers for it. The thesis of this article is that such knowledge does not reside in the model, and that the competitive advantage of a vertical product arises precisely from what is built around the model.
The model is one component, and the most replaceable
The quality of an applied system depends on four elements: the available data and their structure; the rules of the process the system enters; the tools through which the model can act (reading a register, querying an archive, producing a document in the required format); the way the user verifies the result. The language model is only one of these elements and is, by construction, the easiest to replace: vendor interfaces converge, costs fall, same-generation models perform ever closer to one another. What cannot be replaced in an afternoon is everything else.
Embedded knowledge
Domain ontologies, validation rules, workflows and interfaces encapsulate years of sector knowledge in executable form. Knowing that in a cooperative the patronage refund follows rules of its own and is not a dividend; that a transport document has different mandatory fields depending on goods and route; that a balance-sheet item is read together with its explanatory note: this knowledge precedes the model and constrains it. The direction is consistent with what has been observed in specialist domains, where models trained or adapted on the sector's language show advantages on in-domain tasks (in finance, BloombergGPT is the well-known case: Wu et al., 2023); but adapting the model is only one part, and often the lesser one, of adapting the system.
Why the generalist, alone, is not enough
Three gaps are structural. The first concerns the semantics of error: a generalist model does not possess the criterion that separates right from wrong in the domain, and a plausible but wrong answer is indistinguishable to it from a correct one; the notion of error is normative, belongs to the domain, and must be encoded in explicit validations. The second concerns tools: without access to the archives, registers and formats of the process, the model can describe an action but not perform it verifiably. The third concerns acceptability constraints: prescribed formats, deadlines, competences and responsibilities, which decide whether a formally correct output is also usable. All of this can be built around a model, and building it is exactly the work of a vertical product.
Four domains, one method
MurphFin operates on economic and financial data, where the structure of the financial statement is itself knowledge; Trakka on road transport, where documents and their constraints constitute the process; CoopSuite on cooperatives, where the sector's own rules define correctness; Cryterio on information, where sources are the product. The problems differ, the principle does not: start from the real work and build the intelligence around it, not the other way round.
An advantage that compounds
Every real case that passes through the system improves ontologies, validations and workflows: edge cases become rules, errors become checks, exceptions become executable documentation. It is a slow accumulation, hard to compress and hard to replicate, because it requires prolonged exposure to the domain and not merely computing power. This is why Analytiko™ products are designed as domain infrastructure containing a model, not as a model with an interface around it: in the first kind of system the model is an upgradable component; in the second, the only content.
References: Bommasani R. et al., «On the Opportunities and Risks of Foundation Models», arXiv:2108.07258, 2021 · Wu S. et al., «BloombergGPT: A Large Language Model for Finance», arXiv:2303.17564, 2023.
Analytiko · 2 September 2026