Machine learning in financial services is not just a modelling question: it involves data pipelines, software infrastructure, operational processes and controls. A 2019 survey by the Bank of England and the Financial Conduct Authority (FCA) provides a useful snapshot of how UK firms were putting the technology into practiceâand where implementation raised questions for governance and risk management. The findings are not statistically representative of the whole sector, but they illustrate challenges that matter to technical and operational decision-makers.
Machine learning in financial services: from pilots to live use
Two-thirds of survey respondents said they had live machine learning (ML) applications. In this context, âliveâ meant a system was supporting client interactions, business decisions or transactions. The median firm reported using ML in two business areas, while some had more than ten live applications. Respondents also expected their number of applications to grow, although projects in research or development would not necessarily all reach deployment.
Adoption was not limited to one function. Common applications included anti-money laundering (AML), fraud detection and customer engagement. Other reported uses included credit risk, sales and trading, insurance pricing and underwriting, and claims management. These examples show why ML implementation should be understood as a business capability, not simply a model placed into production: it has to connect to the relevant data, systems and decisions.
The survey found that 56% of the reported applications were live, with the remainder in pre-deployment stages such as experimentation and development. Deployment maturity also varied. A model might support a small part of a business line, expand to a larger share of its activities, or reach full deployment. Each stage brings different integration, monitoring and control needs.
Governance must cover the whole ML pipeline
An ML application is a chain of connected processes. It typically includes data acquisition, feature engineering, model development, performance evaluation, validation and deployment. Decisions at one stage affect the next: unreliable input data can undermine a well-chosen algorithm, while a model that performs acceptably in testing may still need safeguards when it is used in live operations.
Data can be structured, semi-structured or unstructured. Structured data, such as information held in a financial database, has clearly defined fields. Semi-structured sources include formats with some organisation, such as website data. Unstructured material, including text and images, has no fixed data model and must be interpreted before it can be used. The survey reported that structured data appeared in more than 80% of use cases, while semi-structured or unstructured data featured in more than two-thirds. Firms often combined data types.
That wider range of inputs makes data quality and oversight particularly important. Teams need to consider whether data is accurate, missing, biased or suitable for the intended purpose. Feature engineeringâthe process of selecting, transforming or combining variablesâalso needs scrutiny, since changes made before modelling can affect what a system learns.
Model validation continues after deployment
Model engineering is iterative: teams select and train models, assess performance, and may revisit their features or algorithms. The appropriate approach depends on the task and available data. The survey identified tree-based models, natural language processing and neural networks among the methods firms reported using. More complex models can be harder to interpret than conventional linear approaches, making it important to test not only whether a model produces useful outputs, but also whether its behaviour is understood sufficiently for its intended use.
Respondents commonly used outcome-focused monitoring and comparisons with benchmarks to assess performance. Other approaches included comparing results with a non-ML model, testing inputs and outputs, using explainability tools, reviewing engineered features and checking data quality. Validation can take place before deployment and continue once an application is live. Ongoing checks help identify when performance changes or when the model no longer behaves as expected.
This is especially relevant to model drift: model outcomes can change over time as new or different data is encountered, or as an application is updated. A one-time test cannot establish that a system will remain suitable throughout its operational life. Monitoring arrangements should therefore be linked to clear escalation and response processes.
Practical safeguards and implementation constraints
Many respondents applied their existing model risk management frameworks to ML, while recognising that these frameworks may need to evolve as applications become more mature or sophisticated. The survey also highlighted the importance of staff capability: business, risk, data and technical teams need enough understanding to assess model behaviour and its implications.
Common safeguards included alert systems and human-in-the-loop mechanisms. Depending on the application, other controls described in the report included guardrails, kill switches and backup systems. These measures can help teams detect unusual behaviour or intervene when a model does not work as intended; they do not replace sound development, validation or governance.
Most reported applications were designed and implemented internally, although firms also used third-party models, software and infrastructure. That distinction matters operationally: internal development does not necessarily mean every component is internally built or hosted. Firms may still depend on external libraries, platforms or cloud services, and need to understand how these fit into the application and its controls.
Respondents identified internal obstacles as more significant than regulation. Legacy IT, limited data and difficulty integrating ML into existing business processes were among the constraints most often cited. Most did not consider relevant regulation an unjustified barrier, though some wanted greater clarity on how existing requirements applied to ML. In practice, deployment planning therefore needs to address both technical integration and the organisationâs ability to manage the resulting system.
What engineering and operations teams should take away
The surveyâs central lesson is that ML risk is not necessarily a wholly new category of risk; rather, technical complexity and changing data can amplify familiar concerns around performance, data quality, oversight and customer outcomes. For teams evaluating an application, practical questions include:
- Are the data sources appropriate, understood and subject to quality checks?
- How will performance be assessed before launch and monitored in live operation?
- Can staff identify when the system needs review or human intervention?
- How do third-party components, infrastructure and internal processes fit into the control framework?
- Are governance arrangements proportionate to the applicationâs purpose, scale and complexity?
For plant managers and automation professionals, the parallel is the need to treat an intelligent application as part of an operational system, not as an isolated algorithm. Clear interfaces, defined responsibilities, monitoring and a planned response to unexpected behaviour all matter. The 2019 findings offer a historical snapshot rather than a current measure of adoption, but their practical implication remains clear: responsible deployment depends on managing the entire lifecycleâfrom data and development through validation, integration and ongoing oversight.
Access Official Technical PDF