Felix Van De Maele

The through-line

Across every source, Van De Maele returns to one claim: the constraint on enterprise AI is no longer the model, it is what surrounds it. In mid-2026 he starts framing this explicitly as a category shift, "for years, enterprise AI was defined by model capability. Today, the models are no longer the limiting factor" , and repeats variations of "the failure point in enterprise AI is rarely the model. It is the context around it" . Over the course of these posts the argument sharpens from a diagnosis into a prescription: first he names the gap (context that exists but can't be reached by agents) , then he ships a product answer to it (the Governed Context Compiler) , then he extends the argument into unstructured data via acquisition , and finally he reframes the whole category around enforcement, ROI, and budget allocation rather than model selection . The throughline hardens from "context matters" to "governed context that flows is the product," with hallucination reframed not as a technical glitch but as a business cost he calls the hallucination tax .

On governed context

Van De Maele's central distinction is between context that merely exists and context that is governed and delivered. "More context isn't the win. Governed context is" . He argues catalogued definitions are inert unless agents can act on them: "A definition an agent can't access is just documentation. Governed context can't simply exist. It has to flow" . This is why two AI applications can give different answers to the same question about monthly active users, "neither model is wrong. They are working from different business definitions" . His fix is architectural: pulling "governed context from the Collibra Knowledge Graph" and delivering it "through REST APIs, MCP, or YAML export," built to be "deterministic by design" . He extends this to the semantic layer: "A semantic layer gives AI a shared business language. Context governance builds on that foundation by providing the meaning, trust, lineage, and policy AI needs to reason reliably and act with confidence" .

On the hallucination tax

He coins and repeats a specific cost concept: "Every time your team has to second-guess an output, manually verify a recommendation, or redo an AI-driven task, you're paying the 'hallucination tax'" . The stakes escalate as AI moves from answering to acting: "poor governance isn't just an accuracy problem. It's a business liability" . He uses a concrete benchmark to make the point vivid, "if a company like Anthropic only achieved a 21% accuracy rate on its own data without the right business context, what do you think happens inside the average enterprise?" . Elsewhere he frames the tax as an emotional as much as operational cost: "AI confidently producing answers that simply can't be trusted" .

On enforcement and trust

Van De Maele distinguishes having a policy from being able to enforce it: "Most have an AI policy. Far fewer can tell me which production models are running without a documented accuracy score... That is the gap in AI governance right now. Not the absence of rules, but the ability to enforce them" . He backs this with third-party data points he treats as validating evidence: Forrester found that "out of the thousands of organizations they surveyed, exactly zero reported a 2x return on their AI investment" while "nearly three-quarters plan to increase their AI spending" , and he cites a colleague's formulation that stuck with him, "Data and knowledge governance is the runtime memory of an enterprise agent" . He also leans on Gartner's prediction that "By 2027, 60% of organizations will fail to realize AI value due to a lack of integration between data governance and AI governance" to argue that fragmented tooling "doesn't scale" .

On unstructured data

He treats unstructured content as the next governance frontier rather than a side case. "Every AI model enters your enterprise knowing almost everything about the world, and almost nothing about your business," including whether a policy changed yesterday or what it's allowed to act on . He frames the acquisition of Deasy Labs as extending governance "into the 80% of enterprise data that has traditionally been out of reach: documents, contracts, call transcripts, emails, and other unstructured content that AI agents depend on to reason accurately and act with confidence" . His broader claim is that "the quality of an agent's decisions ultimately depends on something much more fundamental: the context it has access to," and much of that "lives in unstructured data" .

From the stage

In interview and podcast settings he gives sequencing and mechanism detail the written posts skip. On go-to-market, he identifies product-market fit as the necessary first step before scaling "through thought leadership, product leadership, and market leadership in that specific order" ¹⁷. On building the company, he stresses that "hiring the right people, particularly co-founders and the first 10-20 employees, is essential to building company culture and avoiding the need for micromanagement as the company grows" ¹⁷. On the category itself, he describes Collibra's evolution into an "enterprise AI control plane" and explains that effective agentic AI requires "aligning technical and business teams to build semantic models and ontologies that teach agents how to interpret and navigate business processes, not just connecting data to agents" ¹⁸, with the stated aim of helping organizations "run agents faster and cheaper" in production ¹⁸. He also points to the pandemic as a public illustration of the trust problem, noting that COVID-19 dashboards and statistics forced society to confront "questions about data trustworthiness" ¹⁶.

Takeaways

• Don't just catalog business definitions, make sure agents can actually reach them at runtime; an inaccessible definition is "just documentation" .
• Before adding model spend, audit whether production models even have documented accuracy scores or classified training data, since that gap is where governance actually breaks .
• Treat unstructured content (contracts, transcripts, emails) as governance scope, not just structured data, since it's roughly 80% of enterprise data agents depend on .
• Expect near-zero AI ROI without integrated governance: Forrester found zero organizations reporting 2x AI ROI even as spending keeps rising .
• When budgeting for AI, protect the data and governance layer rather than raiding it to fund models and infrastructure .
• If scaling internationally, invest as much in culture and team-building as in go-to-market, and hire founding and early employees deliberately to avoid future micromanagement ¹⁷.