Gold In, Gold Out: Takeaways from EDM Association DataVision EMEA 2026 in London
- Daniel Rolles
- 3 minutes ago
- 8 min read
By Daniel Rolles, BearingNode
A few weeks ago, I had the privilege of attending the EDM Association's DataVision EMEA 2026 conference in London. The theme — Gold In, Gold Out: Building Trusted Data to Optimize AI — set the tone for a day of sharp, practitioner-level conversation about where data management and AI are heading. It was the kind of event where the hallway discussions were as valuable as the stage sessions, and I came away with a notebook full of insights, new connections, and a clearer view of the challenges the industry is grappling with right now.
Here are my key takeaways.
The Big Shift: Data as Capital
The opening keynote from James Dallas, Head of Investment Bank at UBS, framed the day's discussion perfectly. His message was blunt: AI is shining a light on the cracks in our data foundations, and the banks that treat data like capital — with the same rigour, ownership, and measurement — will be the ones that succeed.
A few phrases stuck with me:
• "Never been cheaper to get to a demo, never been more expensive to get to production." The gap between a proof-of-concept and a production-grade AI system remains enormous, and it is overwhelmingly a data problem.
• "Better the brakes, the faster you can go." Good governance enables innovation rather than throttling it. Embed governance into the architecture, not as a process overlay.
• "Lineage is too hard." This came up repeatedly throughout the day. Despite being a perennial pain point, data lineage remains the Holy Grail for financial services — and the industry is still searching.
James also touched on what he characterised as the traditional model of data ownership and the challenge of standardisation. How do you measure the data? That question — deceptively simple — surfaced again and again across every session.
Reusing What You Have: AI Risk Management at LSEG
One of the most practical and level-headed sessions came from Sanja Hukovic, Group Director and Head of Model and AI Risk Management at LSEG, alongside Emma York, Corporate CDO and Head of Enterprise Information Governance at LSEG.
Sanja walked through LSEG's AI Risk Management Framework — a structured approach that defines AI risk across six pillars: Responsible AI Concepts, AI Development, AI Flow and Inventory, AI Risk and Controls, AI Lifecycle, and Roles and Responsibilities. The framework extends existing model tiering into AI tiering, ensuring that validation efforts are commensurate with risk.
What made this session genuinely useful, though, was the emphasis on reusing existing risk frameworks rather than building new ones from scratch. The guidance from the session was explicit:
1. Aim for non-invasive governance by design — governance as a product feature, not a process.
2. Shift from "human-led, tool-supported" to "agent-led, human-in-the-loop."
3. Embed controls where work actually happens.
4. Use AI to enforce governance; don't wait for governance to enable AI.
5. Measure success by friction removed, not controls added.
6. Adjust governance so it's flexible and doesn't prevent innovation.
7. Use what you already have — you don't necessarily need a new policy.
Emma spoke about a programme looking at strategic data assets and introduced the concept of "data passporting" — the idea that data assets carry their provenance, quality, and governance context with them as they move through the organisation. Aga Strandskov, Head of Data Strategy at Digital Isle of Man, added a prescient observation: "machines are the consumers of data." The implications for data quality, metadata, and lineage are profound when your end consumer is an AI agent rather than a human analyst.
The LSEG team also shared the CDMC for AI, Data & Analytics Controls (ADAC) framework — a set of specialised capabilities layered on top of the existing Cloud Data Management Capabilities (CDMC) standard. ADAC isn't just a framework; it's an operational approach that maps specific controls (ethics and privacy impact assessments, performance evaluations, model risk management) onto shared data and analytics capabilities (cataloguing, lineage, quality measurement, entitlements) and foundational data management practices.
Data Products: Supply, Demand, and the Measurement Problem
The afternoon panel on data products — featuring Pablo Kotey, Head of Data Enablement at Schroders; Ben Clinch, Data and AI Architecture Advocate; Rob Wentz, Senior Advisor at EDM Association; Nikhil Rao, Senior Expert at McKinsey Digital; and Tina Salvage, Data and AI Governance Consultant at OMNIADIGITAL — was a highlight.
Pablo's contribution resonated strongly with me. He spoke about data products and product-market fit — and crucially, about tracking usage. This is something I've been advocating for some time: if we're building data products, we need to measure whether anyone is actually consuming them, and whether they're delivering value.
Two themes emerged:
• Supply-side solutions alone aren't enough. We need to understand and articulate the data demand stack. Who is asking for what, and why?
• Measurement is the gap. One example that came up was the idea of a CFO tracking the ratio of data products serving each business process as a metric in their quarterly business review. That kind of top-down accountability is still rare. The question of token burn — are the gains from AI being measured against the computational cost? — also surfaced, and it's one I think will become increasingly urgent.
There was a notable observation: no real progress on open standards for data products was reported. The Data Product Standard (DPROD) from EDM Association is a step in the right direction, but adoption remains fragmented.
CDMC + AI: Key Controls for an AI World
Oli Bage, who led the development of the CDMC standard, delivered a spotlight session on how CDMC is being extended to address AI and analytics use cases. The framework now defines three control layers:
• AI Use-Case Governance (specialised): AI use-case governance, ethics and privacy impact analysis, performance evaluations, model risk management.
• Data and Analytics (shared): enhanced audit, functional lineage, transparency, versioning, and reproducibility.
• Data (foundational): cataloguing, ownership, classification, data quality, data lineage, entitlements, and access tracking.
The practical guidance was clear: there is a suggested order of implementation, with certain controls required during development, others always required with real data, and some required before production. This is a maturity-model approach — and it maps closely to how we think about staged capability development in our own framework work.
Agentic AI: New Risks, Old Patterns
The session on agentic AI risks was eye-opening. The risks outlined — unintended consequences, cascading failures, hallucinations, security vulnerabilities, vendor lock-in, ethical and legal compliance boundaries, shadow adoption, complexity overload, resilience and resource continuity — read like a checklist for the next decade of AI governance challenges.
What struck me was how many of these risks map onto patterns we've seen before in data management. The cascading risk — where initial biases or errors are magnified through decision-making processes — is a data quality problem in AI clothing. The vendor lock-in risk is the same conversation we've been having about data platform vendors for years. The resilience concern — automated workflows dependent on agents without defined failover capabilities — is a business continuity problem reframed for the AI era.
BCBS 239 and Risk Data: Back from the Dead
Stella Cabrera, Senior Advisor at EDM Association; Scott Beange, Head of Data Management and Cyber Strategy at Projective Group; and Dominic Sanchez, Global Data Programs Delivery Director at SMBC Group led a session on reigniting BCBS 239 in the age of AI. The message was clear: the principles of BCBS 239 — data quality, lineage, governance, and reporting — are more relevant than ever, not less, precisely because AI amplifies both the value and the risk of poor data foundations.
The Data Excellence Gym: Coaching Your Business to AI Fitness
The closing panel — Roberto Maranca, Energy Management Data Officer at Schneider Electric; Paulomi Shah, Head of Cross Domain Services, Data Glossary and Taxonomy Standards at UBS; Sholthana Begum, Senior Director Technology and Digital at the Competition and Markets Authority; and Kyle Morton, Chief Operating Officer — brought the day's themes together under a fitness metaphor. The message: data excellence isn't a project, it's a regime. You don't get fit once; you stay fit through continuous discipline, measurement, and coaching.
That metaphor works because it captures something the financial services industry has been slow to accept: data management is not a transformation programme with an end date. It is an operational capability that must be sustained, measured, and continuously improved — just like risk management, just like security, just like any other critical business function.
My Three Big Takeaway Themes
Sitting on the train back from London and reviewing my notes, three themes kept surfacing:
1. Governance Does Not Equal Observability
This is a drum I've been beating for a while, and DataVision confirmed it's a live conversation across the industry. There is a constant misunderstanding between governance, management, and observability — and now we're adding AI governance, AI management, and AI observability into the mix. These are different things:
• Governance defines the policies, standards, and accountabilities.
• Management is the operational execution against those policies.
• Observability is the real-time, evidence-based capability to see what's actually happening in your data and AI systems.
You can have excellent governance and still be flying blind operationally. You can have great observability tooling and still lack the governance framework to act on what you see. They need each other, but they are not the same — and vendors routinely conflate them.
2. Open Standards Are Gaining Traction — Finally
Throughout the day, the conversation around open standards — OpenLineage, OpenTelemetry, and the broader push toward reducing vendor lock-in — was the most animated I've heard at any industry event. Large financial institutions are now actively thinking about how to adhere to and be compatible with open standards.
Let me be clear about our position: we are not anti-vendor or anti-commercial software. Vendors do great work. What we are is pro-open-standards and pro the reduction of switching costs and lock-in. That's a material shift in the market — institutions that were previously happy to buy monolithic platforms are now asking hard questions about portability, interoperability, and exit strategies.
3. Data for AI, AI for Data
This concept was touched on but not nearly explored enough: we are simultaneously dealing with AI as a consumer of data (machines consuming data on behalf of humans, agents making decisions based on data pipelines) and AI as a producer and enabler of better data management (using AI agents to accelerate data quality measurement, automate lineage discovery, and manage metadata at scale).
The phrase that captures this — "Data for AI, AI for Data" — is more than a neat slogan. It describes a feedback loop that will define the next phase of data management maturity. The unstructured-versus-structured data debate is still unresolved, the challenge of measuring evals for generative AI outputs is still nascent, and data lineage continues to be universally acknowledged as essential and universally acknowledged as hard.
What's Next
The conversations at DataVision EMEA 2026 validated a lot of what we've been building at BearingNode — particularly the focus on operationalising governance through observability, the importance of open standards, and the need for practical, staged approaches to data and AI maturity. The BearingNode Knowledge Graph (BNKG), with its structured approach to capabilities, sub-capabilities, requirements, and design decisions, maps directly onto many of the challenges aired at the event.
The ADAC working group meetings are something we'll be tracking closely. The controls and mappings being developed there — covering AI use-case governance, ethics and privacy impact assessments, model risk management, and lineage transparency — are exactly the kind of structured, reusable knowledge that belongs in a knowledge graph.
If the days of "garbage in, garbage out" are truly over, then the age of "gold in, gold out" demands more than aspiration. It demands the frameworks, the tooling, and the operational discipline to make trusted data a reality. DataVision EMEA 2026 showed that the industry is ready for that conversation. The question now is who will execute on it.
Daniel Rolles is a co-founder at BearingNode, where he focuses on data and AI observability, governance, and operational maturity. BearingNode helps financial services organisations build the data foundations needed to scale AI responsibly — using open standards, structured frameworks, and the BearingNode Knowledge Graph (BNKG).
About EDM Association DataVision EMEA
DataVision EMEA is the EDM Association's flagship European conference, bringing together data and analytics leaders from financial services and regulated industries to explore practical approaches to data management, governance, and AI enablement. The 2026 edition focused on building trusted data foundations for AI adoption.
Learn more: EDM Association
