Provocation
Palantir solves the ontology, and the layer beneath it is still your work
Palantir's ontology sits on datasets. The work of getting data in, transforming it and keeping it canonical stays in the layer below.
On a recent call, the data and AI go-to-market lead at a systems integrator put the split plainly. He was describing Palantir.
"They are not solving the data ingestion transformation problem, right? They are solving the ontology problem."
Read the quote as a map. The ontology is the layer that gives objects, links and actions to the business, and Palantir is good at it. The layer below is where data gets in, gets shaped, gets tested and stays canonical. Your data engineering agent works in that layer, on your data platform, inside your estate's rules.
The distinction matters because the ontology is the visible part of the stack. It is the part a buyer can see in a demo and write into a business case. The work beneath it is harder to see and more expensive to keep.
The ontology is a layer above the datasets it describes
Palantir's documentation is direct about where the Ontology sits. It is "an operational layer for the organization" that "sits on top of the digital assets integrated into the Palantir platform (datasets, virtual tables, and models)" (Palantir Ontology overview).
That design is coherent. The ontology maps those assets to objects, properties and links, then adds actions and functions so people and applications can act on them. It turns a set of tables into a model of orders, equipment, customers or whatever the business calls its world.
A model of the world is still a model, and it describes only the data it is given. If the dataset behind an object is late, incomplete or shaped at the wrong grain, the object carries the same defect with a more confident label.
The ontology is the top of the stack. The stack is the thing that has to be true.
The datasets under the ontology are made by work the ontology does not do
Objects in the ontology are mapped from existing datasources, whether a dataset, a virtual table or a model. Each one came from somewhere: a source system, a file drop, an API or another team's product. Someone connected it, landed it, typed it and wrote the transformation that gives it a grain.
That work is estate-specific. The source schema is whatever the vendor shipped, and the history is whatever the organisation accumulated. The definitions are whatever finance and operations agreed and then changed. A column added by a vendor, or a region that redefines margin, changes what the ontology shows without an ontology edit.
The ontology gives those definitions a shape. Choosing them and testing the pipeline that produces their data remain separate jobs.
This is the part practitioners already know. A semantic layer can define revenue correctly and still serve revenue from a table that dropped a region. The definition is right, and the data is wrong. The ontology then presents the wrong number with the right name.
That failure belongs to the layer below. The ontology carries the truth of the data it is given, and someone has to own that data.
| Question | The ontology | The data-engineering layer |
|---|---|---|
| What is this thing? | Object types, properties and links | The table, grain and keys behind them |
| What can I do with it? | Actions and functions | The transformation and the test that allow it |
| Is it current? | The state the platform last indexed | Freshness, schedule, failure handling and data quality |
| Who answers when it is wrong? | The application or model owner | The pipeline owner, the contract and the test |
| What changes it? | An ontology edit | A source change, a definition change, a new consumer |
An end-to-end deployment can cover the lower layers for a moment
The counter-position deserves its full weight. Palantir's platform includes data connectivity, integration and pipeline building alongside the Ontology (Palantir data integration). An end-to-end deployment can connect the sources, build the transforms and stand up the ontology in one programme.
A buyer can reasonably conclude that the lower layers are covered. The programme delivered them. The ontology is what remains, so the ontology becomes the work.
That conclusion describes the platform accurately and misses the handover, because a deployment covers a scope. It connects the sources in the room, builds the pipelines in the backlog and defines the objects in the business case. The estate it leaves behind is a snapshot with a maintenance bill.
The bill comes due the first time a source adds a column, a team changes a definition, or a new consumer asks a question the ontology was not modelled for. Nothing in the deployment schedule assigns that work to the ontology.
The deployments this argument concerns also have a commercial shape. The programme ends, while the platform subscription continues and the estate keeps producing new sources and new definitions, whoever is on the other side of the handover. That handover is where a project becomes a capability, and the capability has to live somewhere.
A platform that includes the lower layers still hands back the work
The platform can supply the tools for the lower layers without taking on the ownership of them. Someone still decides that a source is authoritative, that a grain is correct, that a test is the right test and that a definition has changed.
That decision is a data-engineering decision. It needs the pipeline, the model, the data slice and the tests in one place, with a change process that can prove the result.
Palantir supplies parts of that in its platform, while your team supplies the knowledge of the estate, the business rules and the judgement about what correct means. The split does not disappear because both layers carry the same vendor's name.
A demo shows the ontology; the estate shows the layer
An ontology demo is a clean object graph. The orders connect to the customers, the equipment connects to the sites, and every property has a name the business recognises. The demo is true for the data loaded for the demo, which can be rebuilt from a clean extract, while an estate cannot be paused.
The estate is messier. A source retired a column without telling anyone, a regional team defines margin one way and the group defines it another, and a feed arrives late on the first day of the month. A key changes when a system is upgraded. None of that is visible in the object graph until it breaks something.
The layer below is where those conditions are handled. Someone profiles the source, writes the transform, adds the test and decides which definition wins. The ontology then shows the result, and the result is only as good as the handling.
That is why a buyer should look past the demo's object graph to the team that will keep it true. A graph stays useful only while that team can absorb the next source change, the next definition change and the next consumer. The demo data is the easy case.
The work returns when a source, a schema or a definition changes
The lower layer is a loop, and it keeps turning. A source schema changes, a transformation breaks, a definition moves because the business reorganised a region, and a test catches the change. Someone fixes it, the ontology reflects the new state, and then the next source arrives.
Each turn of that loop is small, and the total is a team's capacity. The buyer feels the total as delivery throughput and risk, and the practitioner feels it as time to PR and time per PR.
This is where an agent changes the economics. A data engineering agent can run inside the same loop: profile the data, propose the change, run it in isolation, test it against a representative slice and bring back the evidence. The team reviews an argued change instead of assembling one by hand.
The result is the productivity claim: the same team ships more with Vibedata, because more of its capacity goes to accepted changes and less to the assembly around them.
The layer beneath the ontology is where the buyer's risk lives
The buyer's risk is that the ontology is right and the data beneath it has stopped matching the business. A dashboard shows a plausible number, an agent answers with confidence, and the correction lands in a meeting instead of a pipeline.
That gap is expensive to find late. The team spends its capacity explaining which number is right, while the business waits for an answer it can use. The cost lands on the buyer as risk and on the team as arbitration.
A durable lower layer reduces the risk. The pipeline has an owner, the transformation has a test, the definition has a contract, and the change has evidence. The ontology can then describe a state the team can defend.
That is the layer Vibedata works in. It runs on the data platform the customer already uses, isolates every change from production, and keeps the pipeline, the data model, the semantic model and the contract in one change process. The layer under the ontology stays true after the deployment team leaves.
A durable lower layer is four owned commitments
The capabilities are ordinary, and that is the point. They are the parts a deployment can leave behind and a business case can price.
- A named owner for every pipeline that fills an ontology object.
- A test that proves the grain, the freshness promise and the business rule behind each object.
- One change process for the definition, the transformation and the contract, so a change lands in all three.
- An isolated run and evidence for every change before it reaches production.
Each one asks for a team with the capacity to keep the layer current, and none of them needs a new ontology.
Ask who owns the pipeline under the ontology
The next time an ontology demo makes the data layer look solved, ask a narrower question. Who owns the pipeline that fills the object? Who wrote the test that proves the grain? Who decides when a definition changes?
If the answer is the vendor's deployment team, the work is covered for now. If the answer is your team, the layer is still yours, and the ontology has made it visible.
Ask that question before the demo ends, and write the answer into the business case next to the licence cost.
References
- Palantir, Ontology overview.
- Palantir, Data connectivity & integration.
Get the weekly roundup by email
One email a week: what got easier in Vibedata, a recipe to try yourself, and where data engineering is heading.
Or follow the RSS feed.
