Why the Next Era of Enterprise AI May Be Private
For the past few years, adopting AI has largely meant accessing powerful models through services like ChatGPT, Claude, or other external providers. This approach has made advanced AI widely accessible, allowing companies to experiment without building their own infrastructure. But as AI becomes part of everyday business operations, organizations are starting to ask a different question: should they continue relying on external services, or invest in AI capabilities they can control?
In a recent episode of Commit & Push, Damien Filiatrault speaks with Ivan Lee, founder of Datasaur, about this shift toward private AI. Drawing on his experience at Apple and Yahoo, Ivan explains why more organizations are exploring ways to run models within their own infrastructure, build AI agents around existing business processes, and retain control over the knowledge their systems develop.
Why Businesses Are Reconsidering Their AI Infrastructure
For organizations handling sensitive information, using external AI services isn’t always straightforward. Healthcare providers, financial institutions, and other regulated businesses often face strict requirements around how data is stored, processed, and accessed. Even companies without regulatory constraints may be uncomfortable sending proprietary information or internal business data to third-party systems.
Private AI offers an alternative. Instead of relying entirely on externally hosted models, organizations can run AI systems within infrastructure they control, whether that’s their own cloud environment or dedicated servers. This gives them more flexibility over how information is processed and which technologies are used.
Ivan explains that Datasaur initially focused on building data annotation tools before expanding into private AI solutions. As generative AI developed, the company saw growing demand from organizations that wanted access to these capabilities without compromising their data requirements.
But the motivation is changing. Privacy remains important, yet companies are also becoming more interested in controlling costs, customizing their AI systems, and reducing their dependence on outside providers.
The Real Opportunity Is in AI Agents, Not Chatbots
One of Ivan’s more interesting observations is that chatbots may be a temporary stage in how people interact with AI.
He compares today’s chatbot experience to early spellchecking software, when users had to copy text into a separate program, correct it, and paste it back into their documents. Eventually, spellchecking became integrated into the applications people already used. Ivan believes AI is following a similar path, moving from standalone interfaces toward capabilities embedded directly into everyday workflows.
This becomes particularly valuable with AI agents. Rather than requiring employees to manually prompt a chatbot, agents can recognize when work needs to be done, execute routine tasks, and request human input when necessary.
Ivan shares an example from Datasaur’s work with an organization processing Medicare insurance claims. The original workflow involved 133 steps and took approximately 65 hours to complete. By introducing AI agents to handle repetitive activities and coordinate the process, the team reduced that time to around five hours.
Importantly, the system still includes human checkpoints for critical decisions. Automating every step might sound appealing, but a mistake early in a complex workflow can affect everything that follows. The objective isn’t necessarily to remove people from the process, but to let them focus on decisions that require their judgment.
Building Better AI Doesn’t Always Mean Using Bigger Models
The insurance claims example also illustrates another important lesson: not every task needs a powerful language model.
A workflow might involve reading documents, validating information, classifying requests, and coordinating approvals. Some of these activities benefit from an LLM, while others can be handled more efficiently by traditional software or smaller, specialized models.
Using a large model for every operation can introduce unnecessary costs and complexity, particularly when processing thousands or millions of documents. Instead, Ivan describes an approach where different technologies work together, with an LLM potentially coordinating the overall workflow while simpler tools handle individual tasks.
The same practical thinking applies to evaluating AI systems. A model’s performance on public benchmarks doesn’t necessarily indicate how well it will perform a specific business function. Datasaur therefore builds custom evaluations using real examples from each customer’s workflows, allowing the team to measure accuracy against the requirements that actually matter.
This approach also makes it easier to determine whether a newer model is genuinely an improvement. For an organization with a reliable system already in place, upgrading to the latest model may offer little benefit if the existing one continues to perform well.
From Renting AI to Owning Intelligence
Toward the end of the conversation, Ivan raises a broader question about how companies should think about their AI investments.
For years, accessing frontier models through external providers made sense. These services offered capabilities that were difficult to replicate, without requiring organizations to make substantial infrastructure investments. But as open-weight models become more capable and AI usage expands, the economics are beginning to change.
Companies running large numbers of AI-powered workflows may find that usage-based costs accumulate quickly. Hosting their own models introduces infrastructure and maintenance expenses, but it can also provide greater predictability and control.
Ivan argues that the bigger opportunity goes beyond reducing costs. As organizations develop AI systems, they accumulate valuable knowledge through custom evaluations, training examples, workflow definitions, and feedback from real-world use. Retaining those assets internally allows businesses to continue improving systems around their specific requirements.
This doesn’t mean every company needs to operate its own AI infrastructure. External providers will remain a practical choice for many applications. But organizations with sensitive data, substantial AI workloads, or specialized processes have more reasons to consider alternatives.
Ultimately, the next phase of enterprise AI may be less about accessing the most powerful models and more about building systems that work reliably within an organization’s existing operations.
The companies that benefit most from AI may not be those adopting every new model or tool. They may be the ones that understand where AI adds value, integrate it into the right processes, and maintain control over the capabilities that matter to their business.