Adrian Gschwend: Building AI Systems for Government and Enterprise

Adrian Gschwend has spent 25 years working with enterprise data — first in cybersecurity, then in knowledge graphs, long before either term was fashionable. 

Today he is the founder of Qlevia AI, a platform that unifies fragmented enterprise data into a single, explainable system. Deployed across Swiss government infrastructure, regulated industries, and major corporations, it tackles a problem hiding in plain sight across most large organizations.

"The limitation is not the model," he says. "It's the data."

From Tables to Relationships

Modern enterprises are built on data — but almost all of it is still stored in tables. CRM systems, ERP systems, and databases rely on structured rows and columns to organize information. While efficient for storage, this approach doesn't reflect how real-world decisions are actually made.

"If you explain something to someone, you don't draw tables," Gschwend says. "You draw circles and connect them. That's the way you would explain something on a whiteboard — and that's exactly the way we store data in a graph."

Decisions rarely come from a single dataset. They require connecting information across multiple systems, contexts, and timeframes. Traditional data structures were not designed for this.

Why AI Fails in the Enterprise

The recent wave of AI has created the perception that models alone can unlock value from enterprise data. In practice, most organizations have built dashboards, integrated LLMs, and seen very little change in how they actually make decisions.

Gschwend recalls a conversation with the chief pilot of a major international airline, responsible for optimizing fuel consumption — one of the largest cost drivers in aviation. The inputs required were straightforward: historical fuel usage correlated with weather conditions at destination airports. 

The pilot had spent an entire year trying to get his IT department to connect the two datasets. They couldn't do it — their systems were built around tabular, siloed views not designed for connected data problems.

"They have hundreds of dashboards," Gschwend says. "Not one of them solves a real operational problem. They use Databricks, Snowflake, Tableau — you name it. And still, the chief pilot is basically guessing how much fuel to load."

The Missing Semantic Layer

At the core of this limitation is the absence of a semantic layer — a system that defines what data actually means. In traditional databases, meaning is implicit. Column names and table structures provide partial context, but they are often inconsistent, ambiguous, and unreadable by machines.

Most AI applications today focus on unstructured data — documents, conversations, text. But that's only a fraction of what enterprises actually run on. The real value sits in structured systems: CRM, ERP, ten years of operational data that no one can easily query. That's what a knowledge graph unlocks.

Systems like Qlevia are designed to operate at this level — connecting data across entire organizations in real time, rather than querying isolated systems.

From Government Infrastructure to Regulatory Compliance

Gschwend recalls one client — a Swiss insurance company — that was rejected by the financial regulator. The requirement: prove that a failure in one data center would not cascade to another. 

The company had tried to demonstrate this using Excel sheets, pulling data from various systems and manually checking for circular dependencies. The regulator rejected it — the data lineage was unclear, the methodology unconvincing.

With Qlevia, the team mapped the company's cloud deployments directly into the knowledge graph, linking each deployment to its corresponding application. 

The system could then answer the regulator's question precisely: are there any dependencies between data center A and data center B? The answer was no — and for the first time, the company could prove it.

"We solved a regulatory problem," Gschwend says. "I didn't sell them a technical solution. I helped them with compliance. That's what they paid for."

Where We Are on the Hype Cycle

Gschwend draws on Gartner's hype cycle to situate where both knowledge graphs and large language models currently stand. 

Knowledge graphs already passed through the peak of inflated expectations and the trough of disillusionment — and are now on the slope of productivity. 

Large language models, in his assessment, are still near the peak.

"Companies will drop some of the efforts they started, because they got overpromised," he says. "And when they ask themselves why it didn't work, they'll find the same answer: they didn't have the semantic layer between their structured data and the model."

Where Value Will Be Created

This shift has implications for how value is distributed across the AI stack. While much of the current focus is on foundational models, Gschwend expects the market to consolidate significantly.

He draws a parallel with the aviation industry. Boeing and Airbus effectively share the global market for commercial aircraft — not because they formed a cartel, but because the capital requirements for building planes are so enormous that the market can only sustain two or three viable producers. 

The same dynamic, he argues, will play out with frontier AI models.

"The cost of API calls goes down, down, down," he says. "The capability goes up. The real value won't be in the models. It will be in what companies do by using them."

How Enterprise AI Adoption Varies by Region

Gschwend has worked with organizations across Europe, the United States, and Southeast Asia — and the differences in adoption are sharper than he expected. 

The US leads, driven by large financial institutions like Bloomberg, JPMorgan Chase, and Morgan Stanley, which have already built dedicated internal knowledge graph teams. 

Europe comes second — industrial companies facing pressure from China have been among the most aggressive adopters. BMW is one he cites directly.

Southeast Asia has been the surprise. Despite Singapore's reputation as a financial hub, Gschwend finds the adoption curve slower than expected. He points to a structural reason: industries that are simply too profitable have no incentive to change.

"Insurance is simply too good a business," he says. "You can be outdated and still make a lot of money. If there's no competitive pressure, there's no urgency to change."

What AI-Native Organizations Will Look Like

As enterprise systems evolve, Gschwend expects a fundamental shift from transactional to event-based architectures. Traditional systems capture what happened, when, and where. Event-based systems capture why — the conditions and relationships surrounding each action.

He gives the example of a fast-moving consumer goods company that discovers a 20% sales decline in Indonesia only at the end of the quarter. By then, it is too late to respond.

"Operational would mean the system tells you while it's happening," he says. "In the past two weeks, Indonesia has started trending down. You can still act."

Interfaces also shift. Instead of navigating dashboards, users interact with systems through natural language. Gschwend is already building this for himself — an agent that reads his calendar and WhatsApp messages to automatically fill in his time reporting system, matching meetings to clients and logging hours without manual input.

"That's where things are going," he says. "You have a WhatsApp conversation with your CRM. The companies that start betting on this now will be far ahead of everyone else."

The Scaling Breakthrough 

Three years ago, Gschwend discovered the work of Hannah Bast — a German professor who in 2008 wrote the route planner for Google Maps. Her open-source database engine, Qlever, offered performance he describes as orders of magnitude beyond anything else available. When he first saw it, he didn't believe it was real. Today, Qlever sits at the core of Qlevia's platform.

"When I presented her work at Roche in Basel, I had 50 of the smartest people in the knowledge graph domain in the room," Gschwend says. "You could literally see their jaws drop."

Asked what he would be most proud of in ten years, his answer is immediate: convincing Hannah Bast that what she built was the missing piece — and making it the foundation of something that matters.