Your AI Problem Is Usually a Knowledge Problem

If your AI project is stalling, the bottleneck is usually access to your own knowledge. Daniel Fallmann, who founded Mindbreeze back in 2005, said it plainly. The model and the AI stack rarely cause the failure. What matters is context and trusted, relevant information, and most companies already own that knowledge. The hard part is getting the AI to reach it.

In this episode, I talked with Daniel Fallmann, CEO and founder of Mindbreeze. He's been in enterprise search since 2005, long before the LLM craze, helping large organizations pull structured and unstructured data from across the company so people can find the right answer. The work used to be called enterprise search. Now it's enterprise AI search. Daniel's point is that AI didn't change his world much. It sped up understanding information. The hard part stayed the same.

Here's the misconception he wanted to clear up. People say AI projects fail because of hallucination or weak models. In his experience, the real story is different. The bottleneck is whether the AI can reach trusted information with the right context. Every company already has the expertise in house. They have subject matter experts. They have decades of processes that work. The trouble is that the knowledge sits in silos, scattered across systems that were built to store and version records rather than hand context to an AI on demand.

I gave him a small example from a talk I did at a trucking conference. You can ask any model the torque spec on a specific engine part. It might give you the right number, or it might make one up, because that detail isn't public. Connect the actual service manuals to the model and you get an accurate answer. Daniel called that spot on. It's a knowledge project before it's an AI project. Some companies need to fix the knowledge problem first, whether or not they ever turn on the AI.

The idea I keep coming back to is what he calls the augmented employee. Say you hand an NDA to your legal team. A good legal expert spots the risky clauses right away, like an unlimited liability cap or terms about orally disclosed information. Daniel's tooling lets you take that expert's judgment and scale it across the whole company, so a computer scientist like him can run a first pass on a contract without pulling a lawyer into every review. The expert still owns the hard calls. The routine checks get spread to everyone.

That scaling only works if access control is airtight. A friend of mine likes to say you should treat any AI surface like another employee. You give a new hire their own account, their own role, and access to what their job needs. You hold back the keys to everything else. Same with an agent. Daniel was firm on this. The agent should see only what the asking person would be allowed to see, running as them, with those rights enforced. His system builds what he calls a hybrid index, mixing index structures, vector search, and a knowledge graph in one place, and access control has to hold across all of it. Once agents start querying each other through MCP and agent-to-agent protocols, bad access spreads fast. So you validate it firsthand.

He was also clear that governance and trust come from experience, and money can't shortcut them. As he put it, you can't say you raised a billion in venture capital so now your governance is the best out there. That's not how it works. Twenty-one years of being the knowledge backbone for Fortune 500 companies is what built that trust. When his customers ask why a decision got made, he can show the audit trail of which sources fed the answer. That beats shrugging and saying you have no clue where the information came from.

On how to start, his advice surprised me a little. Begin with real-world data and your most critical business processes rather than a safe sandbox. The reason is the CFO. At some point someone asks if the spend is a good investment, and you need a real business case to point to. Run the pilot with your actual subject matter experts, and expect them to show you they're still better than the tool. That's a feature. Daniel pushed back hard on the layoffs narrative, calling a lot of it AI washing. Your experts are the strength of the company. You use them to teach the AI, and when the business scales, you tend to need more people. Growth pulls in hires, it doesn't shed them.

At the end of the day, the question stopped being who has AI. Almost everyone does. The question is who can give the AI the best understanding of their business. For a small shop, that might mean writing down the process that lives in one person's head before they take a week off. For a big company, it means connecting the silos with access rules that hold. Either way, fix the knowledge first. The model is the easy part.