Palantir AI: Architecting Data Decisions for 2026

Palantir’s Foundry platform is designed to take raw data and turn it into something you can actually use to make fast, informed decisions. We’re talking about “AI referrals”, systems that give you proactive recommendations and can even kick off automated workflows, moving way past simple dashboards. So, how do you actually architect a system like this to get real results?

1. Data Ingestion and Ontology Definition

An effective AI referral system in Foundry has to start with getting all your data in one place and then carefully defining an ontology. This means connecting Foundry to every relevant data source you have, from structured Oracle Database tables and SQL Server logs to messy, unstructured documents and even real-time sensor feeds. Foundry’s Data Connection tools are built for this, handling the secure, scalable ingestion of all these different data formats.

With the data loaded, your most important job is building the Ontology. This is where you define the real-world objects in your business, like “Customer,” “Product,” or “Incident”, and spell out exactly how they relate to one another. In a fraud detection project, for example, a “Transaction” object would be linked to a “Customer” and a “Merchant,” with properties defined for each, like “Transaction Amount” and “Customer Location.” You do this in Foundry’s Ontology Manager, which is a visual tool where you literally drag and drop your objects, define their specific properties, and then draw the lines connecting them. Without this semantically rich map, your AI models will have no context and won’t be able to generate useful referrals.

Pro Tip: Start with Core Entities

Resist the urge to model your entire business on day one. Start small. Identify the handful of core entities that are essential for your first referral use case, for supply chain, that might just be “Shipment,” “Warehouse,” and “Supplier.” Build that, prove it works, and then expand from there as new needs pop up. I’ve seen too many projects get bogged down for months trying to build a perfect, all-encompassing ontology from the start which is a recipe for delays and rework. Get to immediate analytical value first.

2. Data Transformation and Harmonization

Your raw data is almost never ready for an AI model. That’s where Foundry’s transformation tools come in. Data engineers and analysts use things like Code Workbook and Pipeline Builder to clean, enrich, and harmonize all the messy source data. If you’re trying to predict equipment failures, for instance, you’ll have sensor data coming in with different formats and units from each manufacturer. You’d jump into Code Workbook and use Python or Spark SQL to standardize those measurements, fill in missing values, and join the tables together into something usable.

Or think about retail, where customer purchase data is often siloed between online and in-store systems. To create a single customer view for product referrals, you have to join those tables, dedupe the customer records, and then aggregate their full purchase history. For this kind of work, Pipeline Builder provides a visual, drag-and-drop flow that lets people with less coding expertise build out the logic. You can set up steps to filter nulls, join tables on the Customer ID, and aggregate total spend per customer. The whole point is to produce a clean, consistent dataset that fits the ontology you defined earlier.

Common Mistake: Neglecting Data Quality at Source

Foundry’s transformation tools are powerful, but they aren’t a magic fix for bad data at the source. The old rule about poor input leading to poor output is absolutely true here. You have to invest in upstream data quality by adding validation rules in your source systems or running regular data audits. Foundry is great at helping you find anomalies, but stopping them from ever being created in the first place will save you a ton of work down the line.

3. Machine Learning Model Development

Once your data is clean, harmonized, and structured by the ontology, you can finally start building the ML models that will power your referrals. Foundry supports this with its Carbon framework and Machine Learning Environment. Data scientists will typically use Code Workbook to develop, train, and then deploy their models. This could be a classification model to predict customer churn (which would trigger a retention referral) or maybe a regression model to forecast demand (which would inform inventory referrals).

Let’s say you’re working on a healthcare project to predict which patients are at high risk for readmission. You’d train a classification model using historical patient data, with features like demographics, their previous diagnoses, the length of their hospital stay, and medication adherence. Inside Foundry, the process involves selecting that training dataset, defining your target variable (like “readmitted within 30 days”), and picking an algorithm like a Random Forest Classifier to train. Foundry gives you the tools you need for model evaluation and versioning, plus explainability features that are absolutely necessary for building trust and meeting compliance in regulated fields. After training, the model is registered and deployed to start making predictions on new data.

For other companies looking at integrating AI with existing systems, see how McKinsey is unifying AI compute by 2026.

Pro Tip: Focus on Model Interpretability

For any referral that actually matters, knowing why the model made its recommendation is just as important as the recommendation itself. You have to use Foundry’s explainability features, like SHAP values or feature importance reports, to give people transparency. This is how you build trust with the end-users who have to act on these referrals and it’s also how you get the feedback needed to make the models better. I guarantee that a black-box model, regardless of its accuracy, will get shot down in any real operational setting.

4. Operationalizing Referrals with Workshop

A standalone ML model is useless. It generates no value until you put its insights into the hands of someone who can act on them. That’s what Workshop, Foundry’s app-building environment, is for. Workshop is how you build the custom, interactive applications that get these AI-generated referrals in front of the right people in a way they can actually understand and use.

For example, a bank might have a model that flags suspicious transactions. A Workshop app would present that referral, “Investigate Transaction X”, to a fraud analyst, but it would also pull in all the context from the ontology: the customer’s transaction history, details on the merchant, the transaction amount, and even the model’s confidence score and the main reasons for the flag. The analyst can then take action right there in the app by clicking “Approve,” “Deny,” or “Escalate,” and that decision is written back to Foundry. This creates a feedback loop that helps improve the model later. Because Workshop is a low-code, drag-and-drop tool, you can spin up these kinds of operational apps incredibly fast.

Common Mistake: Over-Automating Without Human Oversight

It’s always tempting to fully automate everything, particularly for high-volume, low-risk referrals, but you have to resist removing the human from the loop on critical decisions. For anything where AI decisions are at risk, the system works best as a co-pilot, not the pilot. It should highlight a potential problem or opportunity, but a human expert should make the final call. This hybrid model keeps accountability clear and catches the weird edge cases that a model will inevitably miss. I’ve personally seen organizations go all-in on automation, only to have to walk it back after a few costly, unexpected results made them appreciate the need for human judgment.

5. Feedback Loops and Continuous Improvement

Getting from data to a decision is a continuous cycle. To be effective, your Palantir referral system must have tight feedback loops built in from the start. Every single decision made from a referral, whether by a person in a Workshop app or by an automated process, is a new piece of data. That decision data has to be fed back into Foundry to enrich the datasets you’re using for model training.

For instance, if your model recommends a marketing campaign for a certain customer, you must track what happens next. Did they convert? Did they engage? That response is direct feedback on how good the referral was. This new data is then used to retrain the model, making it more accurate over time. Foundry’s data governance tools help make sure this feedback is captured and correctly attributed back in the data pipeline. You also have to constantly monitor your models for performance drift by tracking metrics like precision and recall, which will tell you when it’s time to retrain.

This constant iteration, managed with Foundry’s built-in versioning and deployment tools, is what keeps your AI system from becoming stale and irrelevant. Building the first model is the easy part. The real work is maintaining and evolving it. To see more on how this works in practice, read about how AIFA data boosting agent recommendations by 15%.

In the end, using Palantir AI for referrals changes how a company works, shifting the entire organization from being reactive to proactive. By connecting the dots, ingesting data correctly, building a solid ontology, creating explainable models, operationalizing them in apps, and closing the feedback loop, a business can create a real, sustainable advantage.