Building AI models that you can trust

Driven by the dual driver of efficiency and competitive advantage, financial institutions are in a headlong sprint to integrate artificial intelligence. According to Lloyds Bank, half of UK financial institutions plan to increase their AI investment in the next 12 months, with nearly 60% already reporting improved productivity. The belief is widespread that AI is a key driver for future economic growth.

 

The use cases for AI are wide-ranging, from hyper-personalised marketing to sophisticated fraud detection. Yet, amidst this breakneck adoption, a dangerous assumption has taken root: that more data automatically translates to better, more accurate models. This fallacy is setting the stage for a crisis. The true differentiator in the age of AI isn’t data volume, it’s data provenance.

 

The integrity of an AI’s output is inextricably linked to the integrity of its input. A model trained on a mountain of flawed, biased, or poorly understood data won’t just be ineffective; it will be dangerous.

 

Without a disciplined data infrastructure, AI models are being built on fractured foundations, silently perpetuating historical biases and risking a new, automated wave of credit discrimination. The guardrails against this catastrophe are not new; in fact, banks have been struggling to implement them for over a decade.

 

 

BCBS 239 revisited

The 2008 financial crisis exposed the fatal flaw in banking that was a profound inability to understand and aggregate risk data. The response was BCBS 239, a set of 14 principles issued in 2013 mandating robust risk data aggregation and reporting. Its goals were, and remain, deceptively simple: to ensure data is accurate, complete, and traceable.

 

The industry’s track record with BCBS 239 is poor. The initial compliance deadline was 2016. As of 2024, PwC reports that only two out of 31 globally systemically important banks are fully compliant. The primary culprit is the monumental challenge of modernising IT infrastructure and managing complex, global data flows.

 

For years, many banks viewed BCBS 239 as a costly burden. But the emergence of AI has fundamentally changed the stakes. The principles of BCBS 239, data accuracy, completeness, and traceability, are no longer about risk reporting; they are the foundational bedrock for ethical and effective AI.

 

 

From compliance cost to AI immune system

This is where the concept of data lineage transforms from a compliance checkbox into a strategic imperative. Regulators are now explicitly demanding “complete” and “granular” data lineage, providing a map of every data flow from origin to consumption, while basic lineage from a data catalogue is often insufficient for this task.

 

A modern data lineage framework acts as the essential immune system for AI. It prevents toxic data from poisoning critical decisions by providing full visibility into the data’s journey. Where was this data sourced? What transformations has it undergone? Is it fit for purpose?

 

When an AI model is used to assess a credit application, the bank must be able to trace the data points feeding that decision back to their source, ensuring they are not inadvertently incorporating historical biases based on postcode, gender, or other proxies.

 

Consider the consequences of failure. An AI model trained on decades of lending data without understanding the context of past discriminatory practices could learn to reject qualified applicants from certain demographics, effectively automating and scaling bias. The resulting credit discrimination would not only be a reputational disaster but a profound regulatory and ethical failure.

 

 

A blueprint for trusted AI

There is a clear path forward; the same advanced data lineage technology that enables BCBS 239 compliance is the very tool that makes AI trustworthy and effective. By creating a dynamic, visual map of their data ecosystem, banks can achieve the transparency needed for both regulators and data scientists.

 

The benefits of data lineage are evident. At HSBC, a team using an advanced data lineage solution documented and modelled the entire Wholesale Credit Lending book, 2,000 source tables with 80,000+ fields, in under six months, creating a 90% cost saving while achieving the traceability required for both risk management and transformation. This isn’t only a matter of efficiency; it is about building a trustworthy data foundation.

 

The message for financial institutions is urgent. The race to AI must not be run on broken roads. The decade-long struggle with BCBS 239 is no longer a back-office IT problem; it is the frontline defence against algorithmic discrimination.

 

In the not-too-distant future, it is very likely that banks will face a variety of new regulations specifically developed to address concerns about how AI is used. Rather than taking a wait-and-see approach, banks would be well placed to anticipate some of the requirements they’re likely to face.

 

By finally mastering data accuracy, traceability, and completeness, banks can do more than satisfy regulators; they can build an AI future that is not only innovative but also fair, ethical, and built to last. Investing in a modern data lineage framework isn’t a compliance cost; itis an essential investment in the integrity of every decision your AI will ever make. 

 

---

 

Philip Dutton is the CEO of Solidatus

 

Main image courtesy of iStockPhoto.com and Natee127