Palantir Ontology: Turning ML Predictions into Ordinary Business Object Fields

Source: Palantir Foundry official documentation – Ontology › Models in the Ontology (https://www.palantir.com/docs/foundry/ontology/models/)

Why Many AI Models End Up Unused

Data scientists spend months building a store sales forecasting model that performs well in offline tests. The deliverable is a notebook and a slide deck. Store staff cannot read the notebook, and the ERP system has no place to store the prediction. The model sits idle in a model registry. The core problem: the model works but cannot be plugged into frontline business workflows.

Why Put Models into the Ontology?

Organizations want AI to drive business decisions, but deployment is hard and ROI often falls short. Foundry addresses this with three pillars: a reliable unified data foundation, tools to compare and evaluate models, and a way to embed models into business processes for non-technical users. This article covers the last step: deploying a validated model for business consumption.

Two Paths to Bring Models into the Ontology

Path 1: Real-time Invocation (On-demand, Function Mode) – 4 Steps

Import or build your ML model in Foundry.

Expose it as an online service that accepts requests at any time.

Write a wrapper function with two layers:

Inner wrapper: aligns model inputs and outputs.

Outer wrapper (critical): pulls the object's own attributes from the ontology, automatically assembles the features the model needs, runs inference, and translates the result back into business-meaningful terms.

Call this function from no-code apps or simulation tools – click to compute.

Best for: ad-hoc what-if analysis, e.g., "If this store increases inventory by 20%, how does predicted sales change?"

Path 2: Batch Inference (Pre-compute, Store as Object Property)

Run a single batch job that scores every store or device, save the results as a dataset, and bind that dataset to the ontology objects. The prediction becomes a fixed property on the object – open a store card and see the value.

Best for: dashboards, ranking, filtering, alerts – e.g., find all stores with predicted sales plummeting.

Simple distinction: Batch = pre-computed and stored; everyone sees the same version of the numbers. Real-time = computed on demand; each call reflects the latest data.

Benefits of This Approach

Business users understand and use it (business-level explainability)

Build once, reuse everywhere

A single enterprise-wide digital twin

Field Lessons (Practical Pitfalls)

Fundamental shift: replace complex model APIs with business fields

Predictions are estimates, not facts – must differentiate

Don't mix the two modes indiscriminately

Reuse brings efficiency but also blast radius

Wrapper functions are essentially mandatory when AI agents call models

Next Article Preview

The previous three articles repeatedly mentioned objects, properties, links, actions, functions… The next piece, "Core Concepts", will define each and their relationships, plus a comparison table so database-savvy readers can quickly grasp how the ontology differs from relational tables.