If bad data is the problem, is data governance the answer? AI promises to be smarter and faster than humans at an increasing number of tasks. It can detect patterns scientists might miss, process information faster than analysts and increasingly support decisions that once depended almost entirely on human judgement. In banking, that promise is already becoming reality. AI can...
AI promises to be smarter and faster than humans at an increasing number of tasks. It can detect patterns scientists might miss, process information faster than analysts and increasingly support decisions that once depended almost entirely on human judgement. In banking, that promise is already becoming reality. AI can analyse transactions at enormous scale, identify unusual behaviour, support fraud detection and help banks assess complex patterns of risk.
But even the most sophisticated AI has a surprisingly ordinary dependency: the quality of the data behind it. The principle of garbage in, garbage out existed long before generative AI. Feed a system incomplete, outdated, inconsistent, biased or poorly structured data and its results will reflect those weaknesses. AI doesn’t solve the data-quality problem. If anything, it can make the consequences harder to ignore.
As AI adoption in banking accelerates, data quality in banking is becoming increasingly important, alongside the governance of the information feeding these systems. This raises a central question for banks investing in AI: if bad data is the problem, is data governance the answer?
Bad data is hardly a new problem for banks. What is changing is how much banks are now asking AI to do with it. As AI in banking moves into more important processes, poor-quality data can influence decisions at greater speed and scale. The risk is no longer simply that bad data produces a bad report. It can now feed into AI analysis, create a convincing output and potentially influence a business decision.
That distinction matters, particularly in financial services. Incomplete information could affect a credit-scoring system, inconsistent customer records could undermine KYC checks, and outdated transaction data could influence an AI-supported fraud investigation. In these cases, the output may appear sophisticated and authoritative while still being built on an unreliable foundation. The European Central Bank has been clear about this relationship, warning that “poor data inputs will inevitably lead to unreliable results.”
And not every wrong output carries the same consequences. An inaccurate product recommendation may inconvenience a customer. Poor-quality data influencing a credit, fraud or financial-crime decision could have more serious consequences for customers, banks and regulators. As AI takes on a bigger role in these decisions, trustworthy AI starts with trustworthy data.
That warning becomes more significant when considering the scale of adoption. According to the ECB, more than 85% of banks under European banking supervision already use AI, with use cases including areas such as credit scoring and fraud detection. AI in banking is no longer sitting on the sidelines. Banks are rapidly making AI more capable, but that leaves another question: are they improving the data underneath it at the same speed?
For Dutch banks, AI data quality is already moving from a theoretical concern to a practical challenge. KPMG’s State of AI in Banking 2026 identifies data quality as a major bottleneck for scaling AI in Dutch banks. Large financial institutions have accumulated enormous quantities of information across decades of operations, often spread across legacy systems, business units, databases and different generations of technology.
That creates a difficult starting point for AI. Banks may have to work with fragmented and unstructured data, information spread across legacy systems and gaps in the data available to AI applications. A better AI model doesn’t automatically fix those problems. If the data underneath it can’t be trusted, there is a limit to what the model can reliably deliver.
This is one reason AI transformation and data transformation are becoming harder to separate. At ING, the expansion of data-driven technologies such as AI has been accompanied by continued investment in data quality, data governance and data-risk management. The bank has also applied AI and advanced analytics within areas including KYC and transaction monitoring.
ABN AMRO faces a similar challenge. As the bank phases out legacy technology and embeds AI further into the organisation, improving data quality is also part of its broader transformation. ABN AMRO has described investments in data-quality rules and a “golden source” architecture, while stressing that AI use cases depend on the quality of the underlying data.
Together, these examples highlight a wider challenge for AI adoption in financial services. Banks aren’t only investing in better AI. They are also having to strengthen the data behind it. Building a more capable model does not solve poor data quality on its own. Scaling AI also means making sure those systems have information they can rely on.
This is where data governance for AI becomes important. The term can sound abstract, but its purpose is practical. Data governance establishes who owns data, where it came from, what it means, who can access or change it and how its quality is measured. And as banks use AI more widely, there is another question to answer: should this data actually be used for this particular AI use case?
Doing this effectively requires clear ownership, common standards, data lineage, access controls, quality controls and accountability. The European Banking Authority directly connects robust data-governance frameworks with ensuring the accuracy and reliability of data used by AI models. For banks using AI in regulated processes, knowing where information came from and whether it can be trusted isn’t just a technical concern.
However, there is an important distinction to make. Having data governance does not automatically mean having good data. A bank could have detailed policies, data owners and governance frameworks while still dealing with duplicated customer records, missing information, inconsistent definitions or legacy systems. A governance framework doesn’t suddenly clean decades of data.
What it can do is make those problems easier to identify, measure and address. Clear ownership establishes who is responsible for the data. Standards establish what good data should look like. Lineage helps organisations understand where information originated and how it has changed. Quality controls can then identify when data falls below the required standard.
Data governance therefore doesn’t guarantee good data. It creates the framework through which organisations can identify, control and improve it.
But banks also can’t wait until every dataset is perfect before moving forward with AI. For most large organisations, that is unrealistic. KPMG argues that banks cannot afford to wait for perfect data before moving forward with AI. Instead, they need to determine what level of data quality is acceptable for each AI use case and the risk it carries.
This is where fit-for-purpose data becomes more useful than simply aiming for perfect data. Instead of asking whether a dataset is universally “good”, banks can ask whether it is reliable enough for what they are asking AI to do.
The answer will vary between applications. An AI system recommending content to a customer and one contributing to a credit decision do not carry the same consequences when something goes wrong. Their requirements for data quality, accuracy, explainability and oversight should therefore not necessarily be identical either.
That is where data governance for AI has real value. It does not promise flawless information. It gives banks a way to decide what trustworthy data needs to look like for a particular AI use case and the level of risk involved.
There is also a human side to the trustworthy AI conversation. Data governance doesn’t implement itself. Neither do data platforms, quality controls, data-lineage systems or AI architectures. Behind those capabilities are people.
As banks invest further in AI, they don’t only need professionals who can build sophisticated models. They also need people who can build, govern and maintain the data those models depend on. This reinforces the importance of capabilities across Data Engineering, Data Architecture, Data Governance and Data Quality, alongside AI Engineering and Risk & Compliance Technology.
ING’s continued investment in its data capabilities, tooling and governance provides a useful example. Scaling AI requires more than AI specialists. It also requires professionals who understand how technology, data and governance work together inside complex organisations.
That adds another perspective to the conversation about AI replacing technology jobs. Some roles will undoubtedly change as AI becomes more capable. But as organisations rely more heavily on AI, the expertise needed to make those systems reliable may become even more important.
Someone still needs to understand where the data came from, build the pipelines moving it, establish its architecture, monitor its quality and govern how it is used. Organisations also need people who can determine whether that data is suitable for a particular AI application.
The more organisations depend on AI, the more they may depend on the professionals responsible for making its data trustworthy. For banks, trustworthy AI isn’t only a model challenge. It is a data and people challenge too.
AI in banking is becoming more capable and increasingly embedded in critical processes, but the fundamental principle has not changed: AI is only as smart as the data behind it.
So, if bad data is the problem, is data governance the answer? Not completely.
Governance cannot guarantee perfect information. It cannot eliminate decades of legacy data overnight, prevent every human error or guarantee that an AI system will always reach the correct conclusion. What it can give banks is greater visibility and accountability over the data AI is being asked to trust, supported by clear ownership, standards and quality controls.
The goal isn’t to make every piece of data perfect. It is to understand which data can be trusted, for which purpose, under which conditions and with what level of oversight. As AI moves further into critical banking processes, knowing the difference will matter more.
The AI race isn’t only about building smarter models. It’s about building the data foundations, governance and technical expertise required to trust what those models produce.
We help teams build the specialist capability needed for complex programs. Tell us what you are working on.
