Post Detail ImagePost Detail Image
Building Private AI: What Data Ownership Looks Like in Practice
Contents
AI Industry

Building Private AI: What Data Ownership Looks Like in Practice

Private AI is becoming less about making a broad statement about security and more about making careful engineering decisions. In a recent episode of *Commit & Push*, Damien Filiatrault spoke with Ivan Lee, founder of Datasaur, about why some organizations want to run AI models on their own infrastructure, how smaller
by
Datasaur
on
November 11, 2026

Private AI is becoming less about making a broad statement about security and more about making careful engineering decisions. In a recent episode of Commit & Push, Damien Filiatrault spoke with Ivan Lee, founder of Datasaur, about why some organizations want to run AI models on their own infrastructure, how smaller models can be useful, and what it takes to connect AI to real operational work.

The conversation is ultimately about control. Companies want to understand where their data goes, how it is processed, and what happens to it after a model produces an answer. For organizations handling sensitive information, those questions are not abstract. They shape architecture, procurement, compliance, and the day-to-day experience of the people using AI.

Private AI Starts With Data Ownership

Running models on private infrastructure can give an organization more control over its data and its operating environment. That does not mean every company should build and maintain every part of an AI stack itself. It does mean that teams should be able to make an informed decision about where data is stored, which systems can access it, and which workflows are safe to automate.

For many businesses, the value of private AI is therefore not simply that a model is hosted behind a company firewall. The value is the ability to align the model, the data, and the workflow with the organization’s own requirements. A team can decide how information is retained, how access is managed, and how model behavior is evaluated against the work that matters most.

The Insurance Claims Example

Ivan shared an insurance claims example that makes the potential impact concrete. A process that once took 65 hours was reduced to five. The important point is not that AI replaces every part of the process. The important point is that a long, repetitive workflow can be examined step by step, then redesigned so that people spend less time on mechanical work and more time on decisions that require context and judgment.

That kind of improvement comes from connecting AI to an actual business process. It requires understanding the documents involved, the decisions employees make, the exceptions that appear, and the points where a human must remain accountable. The result is not just a faster model response. It is a better operating system for the work itself.

Why Smaller Models Can Be the Right Choice

The discussion also challenged the assumption that bigger models are always better. A smaller model can make more sense when a task has a clear scope, predictable inputs, and a limited set of acceptable outputs. It may be faster, less expensive to run, and easier to deploy in a controlled environment.

A smaller model can also fit more naturally into a private architecture. If the task is classification, extraction, routing, or another focused operation, the most capable general-purpose model may add complexity without adding useful accuracy. Choosing a model is therefore an engineering decision. The right question is not which model has the largest benchmark score. It is which model performs reliably on the organization’s own work while meeting its requirements for cost, speed, privacy, and oversight.

Test AI Against Real Work

One of the clearest lessons from the conversation is the need to test AI against real work. Demonstrations can show what a system might do, but production workflows expose the details that determine whether it is useful. Real data contains ambiguity, incomplete information, inconsistent formats, and unusual cases that are difficult to reproduce in a polished demo.

Testing against real work also creates a better basis for deciding where people should stay involved. Some steps can be automated with confidence. Others need review, escalation, or a final human decision. That boundary should be discovered through evaluation, not assumed in advance.

This is where engineering judgment becomes central. Teams need to know what to automate, what to measure, and where to preserve human control. They need to look beyond a single impressive output and ask whether the system improves the process over time, whether errors are visible, and whether employees can understand and correct what the AI has done.

Building With Judgment

Private AI is not one product choice or one deployment pattern. It is a way of making decisions about data, models, infrastructure, and people together. The strongest implementations begin with a real workflow, identify the parts that create unnecessary effort, and test different levels of automation against the outcomes the business actually needs.

The conversation between Damien Filiatrault and Ivan Lee offers a practical view of that work. Private infrastructure can support data ownership. Smaller models can be the right tool for focused tasks. And meaningful results come from testing systems against reality, not just presenting them in theory.

For organizations exploring AI agents and open-weight models, the path forward is not to automate everything. It is to build carefully, measure honestly, and keep people involved where their judgment adds the most value.

Listen to the full conversation on YouTube, Spotify, or Apple Podcasts.

No items found.
Related post