Stijn Christiaens

The through-line

Across sixteen years of company building and a steady stream of commentary on AI, Stijn Christiaens returns to one idea: the technology is rarely the hard part. The biggest challenges in data and AI work are people problems and cross-functional collaboration rather than technology itself [16], and the speed of technology change is not what determines adoption, it is the speed at which humans and organizations can adjust [19]. That is why he keeps pointing out that organizations repeat on cloud data warehouses the same mistakes they made on-premise [16], and why he finds it "funny because it is true that every new technology cycle organizations have to re-learn the importance of taking care of the data asset" [5].

What has shifted is the vocabulary rather than the argument. The early framing was data governance and the data asset; the current framing is context, meaning and agentic governance. He now treats context as the thing that decides whether agents work at all, quoting approvingly that "Most agentic failures are failures of context" [7], and he pushes governance forward in time: enterprises taking agentic AI to production "need to structure governance before deployment. Retrofitting it after failure is far harder and more expensive than building it in from the start" [1].

On governance as something you build in, not bolt on

The core position is sequencing. Governance structured before deployment beats governance retrofitted after failure, and the difference is measured in cost and difficulty, not in principle [1]. The counterweight to agentic speed is judgment, and he endorses the warning that "The mistake is to confuse speed with judgment" [12]. His picture of what good governance actually looks like is dynamic rather than static: "A well-governed company is a graph of loops running at different speeds" [9], which sits comfortably with his interest in how architecture and data governance get orchestrated at scale in regulated domains such as open finance [14].

Regulation gets the same treatment. Not all regulation slows business; some of it provides guardrails that give business speed [19]. He points to GDPR spinning off more than 200 privacy regulations worldwide as evidence of reach, while conceding that regulation needs iteration, since outputs like cookie-consent dark patterns did not serve the privacy outcome they were meant to produce [19].

On context and the meaning layer

Context is the current preoccupation and he treats it as infrastructure rather than metadata housekeeping. The strategic claim he amplifies is that you cannot rent the meaning layer [15], and the organizational consequence is that "CDOs who can build and manage this infrastructure of context transform their businesses" [13]. He is candid that the work is genuinely difficult, noting what unpacking context feels like [6]. Ontologies sit inside this same question: whether they are hot again or a flash in the pan, "Their fate is in the hands of your agent" [2], which puts the survival of formal meaning structures squarely on whether agents need them to function.

On AI misconceptions and AI-ready data

Two beliefs he considers wrong among business leaders: that AI requires no data investment, and that AI systems are already reliable when they are still an emerging technology [17]. His counter-prediction is that AI will become more powerful and commoditized while requiring increasingly more data to function effectively [16]. "AI ready data" therefore has two halves: the quality requirements for training machine learning models, and the preparation of both structured and unstructured data for AI systems to draw upon [17]. Unstructured data is where the mass sits, at 80 to 90 percent of organizational data, and the route through it is semantic layers that turn documents into knowledge products [17]. He also flags a failure mode intrinsic to the models themselves, drawing on work on the open-ended homogeneity of language models to argue that LLMs can suffer a "hive mind" syndrome when left alone too long [11].

On data as an asset

The recurring instruction is to treat data as a valuable asset rather than exhaust [16], and the recurring test of whether an organization does so is uncomfortably concrete: "I bet you it's easier for you to buy a book online than it is for you to find your own data in your own organization" [19]. He expects the asset framing to become literal, with data moving onto the balance sheet as the Infonomics research predicted [19]. Alongside that, he expects AI to become the interface to data over the next 10 to 15 years, removing the need for SQL or data science skills and normalizing data for everyone [19]. Trustworthy data products are the unit that carries the asset into use [16].

On category creation and pivoting

Collibra did not set out to be a category creation company. It shifted from semantic data integration to data governance around 2009 and 2010, and the category creation opportunity was spotted inside that shift [19]. He treats category creation as a timing question rather than an ambition: you cannot category-create if the category already exists, and today's AI category creators will be followed by category followers within a few years [19].

He is blunt that pivoting is not the elegant ballerina move it is portrayed as. It is weeks to quarters of constructive conflict among founders operating in a fog of war without metrics, and the post-pivot bet is bigger than the pre-pivot one because the runway is shorter [19]. The temperament required is precise: "You have to have strong opinions but loosely held. So have to be a little bit stubborn about your idea... But you also have to be open enough that you're willing to change your mind if the reality hits you in the face that you're going in the wrong direction" [19]. The failure mode is the founder who cannot do the second half, where "even the reality hitting in them in the face cannot change their ways and then they're just, you know, it's like writing into a wall with your eyes open" [19].

On the founder journey

He pushes back hard on how entrepreneurship is sold. "There's this notion that has started to exist in the market almost like a romantic idea about entrepreneurship. I think that's very dangerous, because I don't think entrepreneurship is for everyone. The first years are typically very hard and very uncertain" [19]. He still lands on encouragement: "The romantic notion of the entrepreneur is a little bit of an illusion, but I do recommend start, right? So people start" [19]. Time compresses in a way he thinks should be priced in like dog years: "A day can feel like a week and a week can feel like a month and a month can feel like a year, not because it's boring or because it's hard, but because so much is happening at the same time" [19], which is why starting at 19 is valuable, since you soak up decades of experience before mortgages and families arrive [19]. On himself he is unsentimental: "One of my strengths is, you know, being a little bit relentless. Or in other words, I maybe don't know how to stop, which is also a dangerous superpower in some ways" [19].

The origin story reinforces the collaboration theme. A move from a local software company into a research lab at the University of Brussels exposed him to multidisciplinary work across NLP, linguistics and computer science, and the actual catalyst was a phone call from someone at a packaging company asking how to solve data integration between systems like ERP and CRM. Four co-founders started Collibra on a friends and family round of 15,000 euros each before seeking outside investors [18]. His first structural warning to founding teams follows from that: be explicitly aligned on what type of company you are building, global category creator or regional player, because the strategies are fundamentally different and misalignment on a 10-year commitment causes major conflict [19].

On investors and control

Choose the individual, not the firm brand. The specific partner on your board is the gatekeeper to all the network value, the relationship can last 13 years or more, and stage, geography and strategy fit matter more than the name [19]. Early-stage and late-stage investors are different species: early-stage investors judge on body language and founder dynamics rather than metrics, and can help navigate co-founder conflicts that late-stage investors do not even recognize as problems [19]. On capital decisions, the founder keeps the gas pedal. Only the founder at the steering wheel can decide whether to inject more capital when they see a straight stretch of road ahead, because the investor in the passenger seat cannot make that determination [19].

On selling and measuring engineering

Enterprise selling is a listening exercise. Avoid demo barfing all 99 features, identify influencers, decision makers and gatekeepers, and understand both organizational and individual needs, including career motivations, before you demo [19]. The same human framing applies to metrics generally: "It's not just about numbers, right? It's about people who do things that result in certain numbers" [19]. What you measure in engineering depends on stage. Early on, measure flexibility and agility to test features toward product-market fit; with first customers, measure ease of install and time-to-live; at scale, measure security, reliability, predictability, and whether new features can roll out to the whole customer base [19].

On uneven distribution and learning in public

"The future is already here. It's just unevenly distributed... Estonia is more in the future than Belgium is" [19]. The unevenness is organizational absorption rather than technological availability [19], and his practical response is to convene. He hosts and promotes learning sessions on the Agentic SDLC at Collibra Brussels with AI Tinkerers Brussels, addressed equally to those who are fans of it and those fearful of it [4][10], points leaders in data and AI toward events near Lausanne [3], and hosts a book launch for The Digital Leadership Practice Test with Stijn Viaene at Collibra in Gare Maritime, Brussels [8].

Takeaways

• Structure agentic AI governance before deployment; retrofitting it after failure is far harder and more expensive than building it in from the start [1].
• Assume most agentic failures are failures of context, and treat the meaning layer as infrastructure you build rather than something you can rent [7][15].
• Expect unstructured data, 80 to 90 percent of organizational data, to be the real AI readiness problem, addressed through semantic layers that turn documents into knowledge products [17].
• Align the founding team explicitly on global category creator versus regional player before anything else; the strategies differ fundamentally and the commitment runs 10 years [19].
• Treat pivoting as weeks to quarters of constructive conflict in a fog of war with no metrics, and remember the post-pivot bet is bigger because the runway is shorter [19].
• Pick the individual partner, not the firm brand; that person is the gatekeeper to the network and the relationship can run 13 years or more [19].
• Change what you measure in engineering by stage: agility for product-market fit, install ease and time-to-live with first customers, then security, reliability and predictability at scale [19].
• In enterprise sales, stop demo barfing all 99 features and map influencers, decision makers and gatekeepers along with their individual career motivations first [19].