Jos Gheerardyn

The through-line

Across the sources, Gheerardyn keeps returning to one problem: governance for models and AI is scaling faster than the organizational and technical infrastructure meant to handle it, and most institutions are solving the same problems in isolation. Early talk material focuses on the mechanics of that problem, how data quality and multi-governance complexity make automation hard ¹¹¹²¹³. The 2026 posts shift toward its economic and regulatory consequences: whether banks should build or buy model risk management (MRM) platforms as AI coding tools make in-house builds tempting , and whether accountability for AI decisions can be outsourced to the system itself . Running underneath both phases is a conviction that the field is fragmented and needs deliberate connective tissue, which is why he frames "Yields on Tour" as the part of his job he most looks forward to .

On build versus buy

Gheerardyn's recurring argument is that the AI-driven ease of building software has changed the wrong half of the equation. "With AI coding tools, our team can build this faster than ever" is true, he concedes, but "what hasn't changed is everything that comes after go-live" . He locates the real cost not in construction but in upkeep: "the build was never the expensive part. Maintenance is" . Because "frameworks keep changing, and a model risk platform has to keep up with all of them," and "that ability to adapt is not bank-specific," he concludes that a vendor spreading maintenance cost across many banks beats each institution "carrying that cost alone" . He restates the bottom line plainly: "the build option in 2026 has a slightly lower initial cost. But the project risk has gone up, not down" .

On human accountability in AI governance

He treats "Human in the Loop" as necessary but insufficient as currently implemented, citing the UN Secretary-General's line that "Machines can inform, but humans must decide" as the standard he wants operationalized, not just stated . For him this requires "technology to implement runtime controls and a proper risk-based oversight framework for the human" rather than a box-checking exercise . He applies the same logic to a real incident: when insurance chatbots gave clients wrong information, he calls the insurers' response "precies het juiste instinct" because "De aansprakelijkheid verschuift niet naar de bot" . His broader point is that organizations misclassify these systems: "Een chatbot is een model," yet "veel bedrijven hun chatbot als IT of als klantendienst" treat it, "Niet als een model dat onder governance valt" .

On regulation and risk-taking incentives

Gheerardyn argues that regulatory postponement does not pause real-world risk: "de betrokken risico's related to deploying and using AI systems" don't stop just because part of the EU AI Act is delayed, since "the purpose of AI risk management is to manage all risks, not just compliance risk" . He frames regulation's function historically, using the Wolf of Wall Street to explain that "one of the main roles of regulatory guidance is to bring the cost of excessive risk taking forward," so that companies moving fast without AI risk discipline eventually face "the high probability event of being fined due to non-compliance" instead of a low-probability accident .

On community and fragmentation in model risk

He sees model risk management as a field where practitioners are isolated by design: "The ones who do are often solving the same problem at the same time, in a different bank, without ever talking to each other," so "the same problem gets solved from scratch, ten times over, behind closed doors" . His response is a deliberate circuit of round tables and dinners across Paris, Brussels, New York and London, which he says "has become the part of the job I look forward to most" , echoing the same sentiment when he calls the informal exchange of "fine-tuning ideas, debating trends, sharing stories and enjoying food" what makes the Yields journey "precious" .

From the stage

In talk and interview material, Gheerardyn goes into technical territory the written posts don't touch. He explains that automating model validation runs into operational rather than technical obstacles, because different model types demand different standardization approaches that resist alignment across an organization ¹¹. He proposes that multi-governance MRM combined with better domain modeling can cut unit costs and maintenance effort by decomposing redundant model-usage combinations ¹¹. On data, he uses his own roomba losing sensor data, and DeepMind's AlphaZero, to illustrate how incomplete or incorrect data breaks machine learning models and why data quality is a core operating risk rather than a side issue ¹². On the future of MRM, he states that AI should assist rather than replace human decision-making in validation, help surface documentation weaknesses, and support real-time monitoring, arguing that traditional periodic review cycles cannot keep pace with AI systems that are built quickly ¹³.

Takeaways

• Before recommending an MRM build, weigh maintenance cost and framework-adaptation effort, not just initial build speed with AI coding tools .
• Treat any customer-facing AI system, including chatbots, as a model subject to governance, not as IT or customer service tooling .
• Design human-in-the-loop controls as runtime technical enforcement plus risk-based oversight, not as a policy statement .
• Continue AI risk management even when compliance deadlines slip, since operational risk does not pause with regulation .
• Use real-time monitoring instead of periodic review cycles for AI systems that are built and changed quickly ¹³.
• Create structured, recurring cross-institution exchange to counter the redundant, siloed problem-solving common in model risk teams .