Why the semantic layer makes a company interoperable, not distinctive, and where the real opportunity of AI lies
Before an agent can act for a company, someone has to write down what a customer is. The sentence sounds like paperwork. It is the moment a firm decides what is allowed to exist in its world.
You notice it the instant the room goes quiet at the question. Is the customer the one who signed, the one who pays, the one who renews? The parent group, or the person inside it? Every answer rules another one out. The agent that then issues the invoices, checks the contracts, writes the offers acts only on what stands in that one definition. What nobody wrote down is not there for it.
The borrowed word
Ontology is the name of the oldest question in metaphysics: what is, what really exists. The question goes back to Aristotle; the word itself dates to the early seventeenth century. In 1993 the computer scientist Tom Gruber took it and gave it a new, still-canonical meaning: an ontology is an explicit specification of a conceptualization. The question of what is became the question of what is fixed. That single shift is the whole story.
Whoever builds their company’s ontology today is doing metaphysics by administrative means. They do not decide how reality is depicted, but which reality counts.
Why now
The trigger is not philosophy, it is need. Point an AI straight at a company’s raw tables and have it turn a question into a database query, and it reassembles the joins and the metrics anew on every request. Ask twice, get two numbers, and the wrong one arrives as confidently as the right one. For a conversation that is enough. For an agent that acts, it is not.
So a layer in between that holds the meaning: what a metric is, how the tables connect, what is permitted. The most-cited model comes from Palantir, and it is tellingly built. There an ontology is not just a catalogue of objects but a set of permitted actions on them. Not a dictionary. A statute book of what the machine may do.
The direction is wrong
It is sold as progress for the machine: at last the AI understands the company. Look closely and it runs the other way. For an agent to work on a firm, the firm has to translate itself into a language a machine reads. The adjustment does not run from the technology to the business. It runs from the business to the technology.
The machine does not learn the company’s language. The company learns the machine’s.
Interoperable, not distinctive
This is where it gets interesting for owners and investors. These meaning layers are turning into standards. In September 2025 Snowflake, Salesforce and dbt Labs launched an initiative meant to unify exactly this: a vendor-neutral standard so that customer or revenue reads the same across every system. A real gain, and its name is interoperability.
Only interoperability and distinctiveness pull against each other. A standard is the language of the common, and what is common sets nothing apart. A shared ontology can capture a firm only as far as it is like all the others. The obvious objection: a company defines its own objects. True. But in a grammar others have set, and that grammar is built for connection, not for the firm’s own edge. A gravity toward the generic is built in: what is genuinely its own loses the connection advantage and ends up not modelled at all. Whoever plugs in pays with distinctiveness.
The SAP reflex
We have built all this once before. For a generation, adopting software meant fitting the company to the reference process. Whoever took SAP got the workflows of a thousand other firms thrown in; anything of its own, where it did not fit the schema, became an expensive special case or vanished. Standardization was the price of scale. Grow large or stay your own, rarely both.
The shared ontology standard is the same reflex one layer up. Again a reference model is rolled out, this time for meaning. Again the firm adapts to the model, not the model to the firm. Only now it wears the label of progress, because an AI runs on it.
What the map leaves out
The move is not new, only its object is. In 1998 James C. Scott described how states make the world legible in order to tax and govern it: cadastral maps instead of grown field boundaries, fixed surnames, uniform measures. And how this making-legible destroys a certain knowledge, the local, practical, unwritten skill he calls by the Greek word metis. The official sees the map. The farmer knows the ground.
The company ontology is the same act one level down. It makes the business legible for the agent. And a firm’s metis, the quiet judgment of the foreman who knows when the rule does not apply, the exception that never sat in a field, fits into no object table. Legibility is not understanding. It is an amputation dressed as self-knowledge.
You notice it in an uncomfortable property of data: its quality is not fixed, it depends on the purpose. The same cleanly booked revenue figure is gold for the audit and worthless for steering in real time. An ontology that fixes one single meaning fights precisely against that mobility.
The unreadable is the moat
And now the uncomfortable consequence. The model is rented, interchangeable, the same for everyone. The edge sits in the layer around it that compounds with every use: in codified judgment, in the firm’s own routine, in what cannot be transferred to a competitor. That is exactly the un-modellable. That it resists being written into a shared standard is not a flaw. It is the reason it is defensible. And it is the one decision an owner cannot delegate: what counts, in its own house, as a customer, a risk, a good order.
Whoever makes themselves fully legible makes themselves replaceable.
The tailored suit that scales
That almost no one has followed yet is not chance, it is the proof. One survey counts the firms that actually run such knowledge graphs in production: from 26 to 27 percent in eighteen months, while the talk about them exploded. What sells is the easy story of legibility, not the thing behind it. The market follows the loud story, not reality.
And this is exactly where the genuinely new thing is thrown away. What AI shifts is not speed and not cost. It is that the tailored suddenly scales. The trade-off that defined the SAP era, large or your own, dissolves. For the first time a firm’s own cut can be brought to scale without being sanded down.
Whoever sees this uses AI not to fit into a shared model but to scale its own distinctiveness. Not the suit off the rack that everyone wears. The tailored one, now in series and still cut to its own body.
The question, then, is not which standard a company adopts, but whether it uses AI to sand its edges down until they fit the model, or to sharpen them. The one is a suit off the rack. The other is the first one that actually fits.
