Calogero Bonasia

Four Ways of Saying “Customer” (2/5): The Dictionary's Missing Cuts

Part 2 of 5. Part 1 introduced the problem and Frontinus.

To see why connection alone is not enough, it helps to look at how retrieval systems work. They rely on similarity. They turn each text into a numerical vector, a so-called dense representation, and return the documents whose vectors fall close to the query's. The method works better than one might expect, because vectors preserve the regularities with which words occur together, and those regularities reflect real relations of meaning.

For the same reason, this method makes two related concepts hard to separate. Terms that occur in the same contexts get nearby vectors, even when people working in the field keep them carefully apart. Espie cites adverse event and safety signal, which in pharmacovigilance denote related but distinct concepts. She also cites the word account, whose meaning changes depending on whether it comes from the CRM, the ERP or the contract repository. The customer I started from is very likely a case of the same kind, because its four senses appear in the same documents and next to the same words.

To name this difficulty I turn to the semiotician I announced. Umberto Eco discussed two models of semantic competence at length (Eco 1984). The dictionary represents meaning through definitions and hierarchical distinctions, on the model of the Porphyrian tree: a genus, a difference, a species, and then another difference. The encyclopedia understands meaning as the network of knowledge a culture associates with each term, a network without a center and without final boundaries, which Eco compared to a labyrinth. His thesis was that the dictionary, taken seriously, dissolves into the encyclopedia. The differences that should separate species are themselves concepts, which refer to other concepts, without end.

Read through these categories, the two halves of the corporate problem become easy to tell apart. Similarity search is a statistical shadow of the encyclopedia: it records associations finely, the family resemblance between concepts, the contexts in which words meet. What it lacks are the dictionary's cuts, the thresholds beyond which the customer stops being the payer and becomes the installation site.

In a company those cuts have never been written down by anyone. A new hire learns them in a few months, by asking, getting things wrong, being corrected out loud. In effect they learn a local dictionary, the portion of the encyclopedia that holds in that company, with the distinctions that company has chosen to make. An artificial system goes into service without that apprenticeship, and no one corrects its mistakes out loud.

Espie shows the consequences with a working example, adapted from a real case. Before a meeting, a sales manager asks the agent what open commitments the company has with Acme on Project Alpha. The question touches at least three systems: the contract repository, which is authoritative for legal commitments; the project system, which records releases and milestones; the CRM, which holds opportunities and the history of the relationship.

A common MCP implementation exposes one tool per system. The solution is reasonable, because it mirrors the organization and keeps the number of tools down. The choice of which tool to call, however, depends on descriptions written in each system's vocabulary, and the word commitments may correspond to any of the three. The typical result is a query to all three, which returns many loosely typed records. The language model is left to work out what they are, and the answer comes out imprecise.

Espie proposes placing a semantic layer between the question and the tools, that is, a written set of definitions and relations. With this layer, open commitments maps to a defined concept residing in the contract repository; Acme to an identified counterparty; Project Alpha to a specific project. The agent calls a single tool, passes it those concepts as typed parameters and receives a traceable result.

In Lector in fabula Eco calls topic the hypothesis by which a reader decides what a text is about. It is the topic that guides the choice, among a word's many possible senses, of the pertinent one (Eco 1979). The semantic layer performs the same function for the agent: it fixes the subject of the question in advance, and in writing.

This imprecision has a measurable cost. On a single question it goes unnoticed. Across thousands of queries a day it becomes a budget line, made of multiplied calls, large volumes of records to filter, reasoning steps added for the sole purpose of discarding what was not needed. A semantic layer reduces the calls and the steps, sometimes to one.

Espie acknowledges that the semantic layer costs too, and a good deal: the time of domain experts, the governance of the model, maintenance that never ends. Her argument concerns the shape of the two expenses. The engineering remedy, made of filters and instructions added case by case, grows with use. The semantic investment is concentrated at the start and pays more with every system added. The argument is plausible, yet it has no figures. I take from it which side the advantage lies on. How large it is remains to be measured.

Espie's conclusion deserves to be reported almost word for word. Where meaning is not modeled, it is inferred; where it cannot be inferred, it is guessed. The scene recalls a passage in Eco. In The Limits of Interpretation he describes a hermetic semiosis that slides from resemblance to resemblance without finding a point at which to stop, and sets against it the constraints the text itself imposes on the interpreter (Eco 1990). An agent without constraints produces the same effect: it generates plausible readings, and plausibility remains the only criterion that selects them.

The problem therefore concerns the organization more than the models, and Enterprise Knowledge itself acknowledges it. For 22 July 2026 the firm announced a keynote panel at Data Architecture Online moderated by Sara Nash, who leads its semantic engineering and artificial intelligence practice. The title spoke of artificial intelligence's reality check and of why most data architectures are not ready for it. The gaps listed in the announcement are fragmented control planes, missing metadata, unclear ownership, inconsistent meanings from one system to another. All of them concern the house of those who use the models. The announcement also presented data governance as the condition that makes artificial intelligence projects possible: a contrarian thesis for anyone who has seen a project run aground in a committee.

Next: part 3, content, context and who answers for meaning.

Sources for this part

#artificial intelligence #english #organizational intelligence #semantic layer