Felix Van De Maele

Felix Van De Maele's public statements over 2026 form a single, tightly repeated argument: enterprise AI does not fail because models are weak, it fails because the context around them is missing, fragmented, or unreachable. He states it almost identically across posts: "The failure point in enterprise AI is rarely the model. It is the context around it" . His illustration is consistent too, the example of two AI applications giving different answers to the same question about monthly active users, "Neither model is wrong. They are working from different business definitions" . For Van De Maele this is not a modeling problem but an organizational one: businesses have spent years building shared definitions and then locked them away where the systems that need them cannot get to them. His phrase for this is blunt and becomes a kind of thesis statement: "A definition an agent can't access is just documentation" .

From that diagnosis flows his central prescriptive idea, that context must be governed and, critically, must move. "Governed context can't simply exist. It has to flow" . He describes this as the rationale for what Collibra built, a "governed context compiler" (later capitalized as "Governed Context Compiler") that pulls governed context from the Collibra Knowledge Graph and pushes it to downstream systems via REST APIs, MCP, or YAML export, "deterministic by design" . This same logic underlies his framing of Collibra's broader evolution: the company is moving from data governance for analytics toward being what he calls an "enterprise AI control plane," aligning technical and business teams to build semantic models and ontologies so agents can actually interpret business processes rather than just being wired to data ¹⁸. He extends this into partnership language too, arguing "every enterprise will need a control plane for AI. A single place to understand what AI systems exist, what data they rely on, what policies apply, how they're performing, and whether they can be trusted" .

A second recurring theme is that unstructured data, not structured data, is the frontier problem. He repeatedly points out that most of what explains "how a business actually operates" sits in contracts, documents, emails, call transcripts, and knowledge bases, information agents depend on but organizations have never governed . He frames the Deasy Labs acquisition as extending Collibra's governed footprint into "the 80% of enterprise data that has traditionally been out of reach" , and ties this to a growth stat he clearly finds validating, that Collibra's managed data assets grew from 46 million in 2019 to over 2 billion today, a "71% compound annual growth rate" . His conclusion each time is not that scale matters for its own sake but that "what matters even more is the trust behind every one of them" .

Van De Maele is also preoccupied with the gap between AI capability and AI trust, and he leans on external research to make the point rather than just asserting it. Citing a Forrester report, he highlights that zero surveyed organizations reported a 2x return on AI investment even as three-quarters plan to increase spending, concluding "the challenge is no longer the technology" . He borrows a line from a colleague at Data Citizens that he says "has stayed with me": "Data and knowledge governance is the runtime memory of an enterprise agent" . Similarly, he cites a Gartner prediction that 60% of organizations will fail to realize AI value due to disconnected data and AI governance, using it to argue against fragmented tooling: "you're asking teams to assemble trust after the fact. It doesn't scale" . This is a consistent rhetorical pattern for him, take a striking external statistic and use it to reinforce a point he was already making about governance as the actual bottleneck.

He pushes back explicitly on the idea that governance and speed are in tension. "One of the biggest misconceptions in AI is that governance slows innovation. The opposite is increasingly true" . He argues that the fastest-moving AI organizations are the ones with "the strongest foundations: trusted data, clear accountability, and governance that enables action rather than restricts it" , and describes governance as evolving "from a control function into a competitive advantage" . This reframing runs through his commentary on incident response as well, where he argues that alerts about pipeline failures or schema drift are useless without business context to answer which models, processes, or policies are affected, and that "the boundaries between data quality, observability, and governance are beginning to disappear" .

On the broader AI hype question, Van De Maele resists binary takes. "The AI conversation has become surprisingly binary," he says, rejecting both blanket optimism and dismissiveness, and instead arguing leaders must hold two ideas simultaneously, that they are "probably underestimating how much AI will change" their business while "probably overestimating how many problems should be solved with AI" . He names the cost of ignoring this the "hallucination tax," which he defines as "AI confidently producing answers that simply can't be trusted" . He makes a related distinction between output and outcome, arguing that "the biggest challenge in AI is no longer generating an answer. It's completing a task," and that success will ultimately "be measured by what it can reliably do" rather than what it can say .

Across nearly two years of commentary, Van De Maele's thinking shows more consistency than evolution, he arrives early at "context is the differentiator, not the model" and spends subsequent posts extending that idea into new product launches (the AI Command Center , the Governed Context Compiler ), new partnerships (Databricks ), new acquisitions (Deasy Labs ), and new borrowed frameworks (Tom Dejonghe's "context gravity," which he calls "a simple but powerful idea... just as data has gravity, context creates its own pull" ). The practical takeaway he keeps returning listeners and readers to is that organizations already have plenty of definitions, policies, and lineage, their failure is architectural, that context sits inert in catalogs and documentation instead of being compiled, governed, and delivered to every agent and application that needs to act on it in the format it consumes.