The AI Gold Rush for Private Data and the Illusive Promise of Agents

Morten Rand-Hendriksen reports back from the AI Summit in San Francisco with a clear-eyed summary: the generative AI industry is going through a grounding moment. The hype is cooling. The hard problems are becoming impossible to ignore. His two main takeaways from the conference point to where the real work is happening now and where the real friction remains.

The Data Race Moves Behind Corporate Walls

The dominant theme at the summit was not flashy new models or benchmark-beating demos. It was data infrastructure: RAG, RAFT, fine-tuning, graph RAG. The reason is straightforward. Large language models produce convincing language but struggle with real-world specifics, especially anything outside their training data. RAG and its variants solve this by feeding relevant data to the model at query time.

But this creates a new problem. The data that matters most -- corporate records, healthcare information, proprietary business knowledge -- sits behind firewalls and access controls. Morten frames this as the next gold rush: figuring out how to access private data securely, train models on it without leakage, and keep sensitive information contained. It is a genuinely hard problem with regulatory, technical, and trust dimensions all tangled together.

What is actually happening, he argues, is a realignment. The industry is swinging back toward classical AI -- machine learning, embeddings, knowledge graphs -- to support the generative layer. Generative AI is good at language. It is not good at the retrieval, reasoning, and structured knowledge tasks that real enterprise use cases demand. Those require older, proven techniques working underneath.

Agents Remain a Promise, Not a Product

The second takeaway is about AI agents, the idea that an AI system can reason about your intent and take actions on your behalf. The canonical example: tell it you want to go to Spain, and it checks your calendar, finds flights matching your preferences, books them, and updates your schedule.

Morten's observation is that even the people building these tools now acknowledge how far away reliable agents are. The difficulty is not just technical capability (though that remains hard). The deeper problem is control. Giving an agent the power to act means it will act. Defining the boundaries of what it should and should not do, under which circumstances, turns out to be an enormously complex governance problem. The models cannot reason like people, so they cannot exercise the kind of judgment that these tasks actually require.

The Grounding Moment

Morten describes the current state of generative AI as a "realignment to how the world really works." The promise is still real. But the challenges are stacking up, and even the strongest believers in the space are starting to acknowledge their weight. The industry spent the last two years selling the dream. Now it has to do the engineering.

Key Takeaways

• The AI industry's focus is shifting from public training data to secure access to private, corporate, and regulated data -- a harder and more consequential problem.
• RAG and related retrieval techniques are now table stakes, but they are pushing the industry back toward classical AI infrastructure (embeddings, graphs, ML pipelines) rather than pure generative approaches.
• AI agents that act autonomously remain far from production-ready. The control and governance problems are at least as hard as the technical ones.
• The generative AI space is entering a grounding phase where real engineering constraints are replacing speculative hype.