The Ontology Ceiling

On July 8, 2026, Databricks flipped its Genie products to pay-as-you-go billing. Genie Agents, Genie Code, Genie One. One hundred and fifty DBUs of free LLM use a month, and after that the meter runs. That is a small billing note. It is also the sound of the market conceding a point most CIOs are still refusing to hear.

Three weeks earlier, at Data + AI Summit 2026, both Databricks and Snowflake shipped what they call semantic context layers. Snowflake calls it Cortex Sense. Databricks calls it Genie's ontology layer. The pitch was identical from both vendors, which almost never happens in this rivalry, and the number they used to sell it was almost identical too. Snowflake reported that Cortex Sense pushed accuracy on structured context tasks from 24 percent to 86 percent. Same models. Same warehouses. Different layer between them.

Two of the biggest data platforms on earth just told you that your data ceiling was more than three times lower than the ceiling on model quality. And you were paying the model bill anyway.

Seven Percent, And The Bill Is Compounding

In March 2026, Harvard Business Review Analytic Services published "Taming the Complexity of AI Data Readiness" with Cloudera. The number that traveled from that report was seven. Seven percent of enterprises said their data was completely ready for AI. More than a quarter said their data was not very ready or not ready at all. This is the real market for data readiness for AI adoption, and the ratio is not moving.

Cloudera followed with an April 14 survey of 1,550 global AI decision-makers. Almost eight in ten said their AI programs were being held back by how their data was walled off across environments. 73 percent said their organization should prioritize AI data quality more than it currently does. An identical 73 percent said preparing data for AI has been hard.

Publicis Sapient's 2026 Global Enterprise AI Report, published this spring off its own survey of 1,550 AI decision-makers, put an even sharper point on the picture. 73 percent said AI is used regularly or across most business processes. Only 10 percent said AI is core to how their business operates. Read those two numbers next to each other. The activity is broad. The dependence is thin. The reason it stays thin is the same reason so many pilots get abandoned. The data will not carry it.

Gartner has been putting a knife on this too. Their line for 2026 is that organizations will abandon 60 percent of AI projects that lack AI-ready data foundations. Read that number as a floor. The pilots that die are dying at the same coordinate, and the coordinate sits upstream of the model choice.

MIT's Project NANDA published "The GenAI Divide: State of AI in Business 2025" with the figure everyone quoted at each other for a whole quarter. 95 percent of enterprise generative AI pilots produced no measurable P&L return. The report was built from 150 executive interviews, 350 employee surveys, and 300 public deployments. Read the report closely and the same words keep coming back. Weak data foundations. Poor integration with real workflows. Unclear definitions of success. Every diagnosis lands in the same place: the shape of the data.

What "AI-Ready" Actually Means Now

Two years ago, "AI-ready data" meant clean and labeled. In 2026 it means something more specific and much harder.

An agent that fetches a row from your warehouse is asking a question. What does this column mean. What does this table represent. Which source of truth wins when two rows disagree. Which of these customer records point to the same customer. What time was this recorded, and in whose timezone. What was true when this row was written that is no longer true now. Which of these transformations happened, and in what order, and why.

If your warehouse cannot answer those questions with authority, the model answers them for you. It guesses. It picks the plausible reading. It compresses across ambiguities. That is where the reported 52 percent fabrication rate on ungoverned enterprise data comes from. It is where hallucination stops looking like a model defect and starts looking like a data defect wearing a fluent mask.

Snowflake's Cortex Sense delta came from making the schema smarter. The model itself did not change. Cortex Sense gave the retrieval layer a way to disambiguate columns, resolve entities, and refuse queries when the answer would require assumptions the metadata does not support. The model got the same result the prompt engineer used to force by hand, at scale, at machine speed. The 62-point delta was sitting in the warehouse the whole time.

The Two Floors And The Ceiling

There is a floor and a ceiling in every AI system that touches enterprise data.

The floor is retrieval. What can be found, and how fast, and how confidently.

The ceiling is ontology. What is even representable in the shape your data has right now.

You can raise the floor with tools. Vector stores, hybrid indexes, better embeddings, tighter chunkers. Every vendor is racing there because the tooling is legible and priceable. Pinecone, Chroma, Qdrant, Milvus, Weaviate, Snowflake Cortex Search, Databricks Vector Search. Pick one and you buy a floor.

The ceiling is different. You cannot buy your way through the ceiling. You have to change the shape of what you hold. That is where nearly every AI transformation program stops and dies, because it looks like a data project, and data projects are boring, and quarterly budgets do not reward boring.

Why Every Pilot Dies At The Same Spot

Look at the shape of an AI pilot that quietly fails. It usually goes like this.

Someone picks a bounded use case. A support triage bot. A finance question answerer. A sales research assistant. A model gets wired up to a warehouse or a document store. The demo works. The pilot ships. Six weeks in, accuracy hits a plateau around 70 to 80 percent. The team rewrites prompts. The team swaps models. The team adds guardrails. Accuracy moves two points and then stops.

That plateau is the ceiling of the ontology the pilot inherited.

The support bot cannot tell the difference between three products because your ticketing system does not either. The finance answerer keeps grabbing the wrong revenue number because your finance team keeps three revenue numbers that were reconciled in an offline spreadsheet nobody put in the warehouse. The sales research assistant confuses two accounts because your CRM has three duplicate records for the same account with different owners, and the merge rule lived in someone's head.

The model is doing its job. It is running the query against the semantic layer you gave it, and your semantic layer is thin. This is what Databricks was gesturing at when they announced Agent Bricks and reported over 100,000 agents running on the platform, processing a quadrillion tokens a year. The units of leverage in that number sit under the agents and tokens. They are the ontology and lineage those agents ran against. Databricks was telling you that the ones that work run on Unity Catalog, and the ones running against no catalog do not work.

Three Kinds Of Data Debt

Data debt for AI has three faces, and they compound at different rates.

The first is definitional debt. Same word, three meanings. A "customer" in your product database is an account. A "customer" in your billing database is a payer. A "customer" in your NPS survey is a respondent. Every model that reads across the three is silently averaging three ontologies. That is what Snowflake and Databricks are selling a semantic layer to fix, and it is fixable, and it is expensive.

The second is lineage debt. What happened to this number between the source and the report. Which transforms fired, which filters applied, which manual overrides landed in an Airflow DAG that got renamed twice. An agent cannot reason about a number whose provenance is unknown. It will still answer. It will just answer confidently and wrong.

The third is contract debt. Which team owns this table. Who is allowed to break it. What happens downstream when the schema changes. This is the debt that turns a working AI system into a broken one at 3 a.m. on a Thursday, because someone in another team shipped a schema migration that quietly dropped a column, and the agent kept answering.

You can measure your data readiness for AI adoption with one question. If a model got every one of your columns wrong tomorrow, could you tell? If the answer is no, the ceiling is lower than you think.

What The Platform Vendors Are Quietly Telling You

Watch what the vendors are actually shipping this year. The keynotes are noise.

Unstructured and Teradata paired in March 2026 to pipe PDFs, spreadsheets, emails, images, and audio into Teradata's Enterprise Vector Store. Read that announcement as a concession that most of your enterprise knowledge does not live in your warehouse at all, and the ingestion layer is where the value now sits. The processing steps they highlighted were classification, metadata enrichment, sensitive-data detection, and selection at scale. Those are governance steps in a wig. They are calling them data prep because "governance" does not sell as well to a VP of engineering.

Databricks shipped ai_parse_document, ai_extract, and ai_classify as SQL functions in June 2026. That reads like a features release. In practice it is a public admission that the way most enterprises turned unstructured content into structured rows was through a human on Fiverr and a mail-merge template, and that pattern will not survive an agent load.

Snowflake's Cortex Sense is a semantic context layer. Databricks' Genie has an ontology layer. The word "layer" is the tell. The vendors have accepted that no model, however good, can lift performance past what the ontology encodes. So they are selling you the ontology as a product. That is a rational business move by them. It is also a warning to you. The layer they are selling is generic. The one your business needs has to know your business.

The Move That Actually Works

If you take one thing from this piece, take this. Buying an AI product without architecting your data readiness for AI adoption is the same trade as buying a car for a road you have not built. The car will run. It will just not go anywhere useful.

The move is a small number of things done in order, and the order matters.

Start with definitions. Pick the five business concepts that show up in every important decision you make. Customer. Product. Order. Employee. Revenue. Write down what each one means, in prose that a new hire could read once. Decide which system owns each one. Publish the decision internally.

Then move to lineage. For the metrics that appear on your board deck, trace the wire. Which system emitted the raw event, which pipeline transformed it, which report renders it. Where does the transformation logic live. Who can change it. Version it. This is what an agent will need to answer any question about the number without inventing one.

Then move to contracts. For the ten tables that matter most, publish a contract that says what fields exist, what they mean, who owns them, and what happens if they change. Enforce the contract with a CI gate. This is what stops the schema migration at 3 a.m. from destroying the trust an agent has built with your customer support team over three months.

Then, and only then, wire up the model.

Sequence matters because these steps compound. Trying to enforce contracts on tables whose lineage you cannot trace produces a false sense of safety. Publishing definitions before you know which system owns which noun produces a political fight rather than a document. Two months on definitions and lineage, then two months on contracts, then a pilot on a real workload. Every enterprise skips those first four months because they feel like plumbing, and every enterprise that skips them ends up on the wrong side of the seven percent.

If you find yourself skipping the first three steps because they feel slow and the model feels fast, you are the reason the seven percent number stays at seven percent.

The Ceiling Is Where The Business Actually Lives

The reason data readiness is the hardest problem for a CEO to fund is that it does not look like it produces revenue. It looks like a plumbing bill. The 62-point Cortex Sense delta is the answer to that objection. The revenue is sitting in the gap between what your model can do and what your data lets it do, and every quarter you do not close the gap is a quarter your competitor buys the same model and closes it first.

A pilot that stalls at 78 percent accuracy is telling you where your ontology stops. Listening to that signal is the difference between an AI program that graduates and an AI program that becomes the pilot cemetery every enterprise now has.

The 95 percent NANDA figure is a warning about how enterprises fund AI. They are paying for the model layer, which is deflating in cost, and starving the layer under it, which is the layer that carries the whole system.

There is a secondary consequence most boards miss. Vendors are now pricing the ceiling. Databricks moving Genie to pay-as-you-go on July 8 is a price signal. Snowflake's Cortex Sense is a price signal. The layer that used to be free plumbing is now a line item, and the line item is metered per query. A firm that has not architected its own ontology is going to pay a vendor for a generic one, and pay again every time an agent runs against it. The vendor bill is the shape of the ceiling made visible.

The Argument For Architecting, Not Buying

Every off-the-shelf semantic layer, every prebuilt ontology, every vendor-shipped Cortex or Genie or Agent Bricks preset ships with a general definition of "customer" and "revenue" and "order." Your business is specific. That specificity is where the moat lives. It is why your finance team can answer a question the vendor's semantic layer cannot. And the moment you buy the vendor's shape and pour your data into it, the moat evaporates. You have handed the model the same shape your competitor has, minus the parts of your business that made you distinct in the first place.

Architecting your data readiness for AI adoption is the act of taking the shape of your specific business and writing it down in a form a machine can read. It is where your competitive advantage now compounds. The company that owns its ontology owns the ceiling on every AI system that touches it. The company that rents its ontology has already sold the ceiling to whoever will buy the layer next.

Agor AI Advisory does this work. We come in. We look at your five most decision-load-carrying nouns. We look at the lineage you actually have against the lineage you believe you have. We look at the contracts you have not published and the migrations that quietly break things. We design the semantic layer that lets an agent answer questions about your business the way your best analyst would, and we build the governance around it so the layer holds up under load. We do this because the model is now the cheap part, and the layer under it is where the accuracy, the auditability, and the moat all live. If the last four months of vendor announcements have not made this obvious, the next four will.