Thursday, 6 August 2026
The Return of Machine Learning as a Value Add
Why Machine Learning still matters
It is tempting to treat agentic AI, workflows and bots as the destination for every organisation. The reality is more practical. Labelling and predicting from data remain major business needs, and traditional machine learning (ML) is often the strongest way to address them.
For structured data, the use cases are familiar and valuable: deciding whether to offer a loan and at what rate, identifying equipment likely to require an upgrade, or detecting patterns that indicate risk. For unstructured data, machine learning and deep learning also remain effective for tasks such as document classification and the identification of suspicious behaviour in video footage.
Generative AI and agentic AI have expanded what is possible, but they have not replaced the established value of machine learning. The enterprise AI toolkit has grown. It has not narrowed.
Where Machine Learning still wins
Machine Learning remains particularly strong for prediction using tabular data. In these settings, simpler and more explainable models can offer a clearer view of the factors driving an outcome. That interpretability can be as important as raw predictive performance, especially where decisions need to be understood and challenged.
Automation still has a role. AutoML can provide a useful benchmark and may identify a model with meaningfully better accuracy. The aim should not be to choose between code and automation on principle, but to understand the trade-off and select the approach that best serves the use case.
Python also remains fundamental to the data science toolset. Even when code is generated with assistance, the quality assurer needs to understand what is being executed, why it works and how to diagnose it when it does not.
Human judgement remains essential
Feature engineering is one of the clearest examples of where human judgement continues to matter. Combining variables, adjusting distributions, simplifying noisy information and identifying a new predictor all depend on a detailed understanding of the business problem.
Context changes from one use case to the next. A date recorded as ‘190624’, for example, may be interpreted in more than one way. Without the right context, a system cannot reliably determine whether it represents 19 June 2024, 24 June 2019 or simply an integer. Human supervision is what resolves that ambiguity and ensures the data is prepared appropriately.
Different stages of the pipeline therefore demand different levels of autonomy. Loading and profiling data can often be automated. Feature engineering is better treated as human-machine collaboration. Defining the problem and selecting the right data should remain human decisions. The role is not to compete with AI, but to act as the pilot supporting the copilot.
Modernising established ML
In Financial Services, machine learning is not new. Credit-risk models and fraud-detection techniques have been used for years. The challenge is that many still sit in legacy environments, rely on older tooling or operate in a form of pseudo-production, with someone manually running a model against new data on a fixed cycle.
This is where modernisation creates value. Cloud platforms can make established models faster and more secure, while MLOps introduces automation for deployment, monitoring, maintenance and retraining. Monitoring for concept drift helps teams understand when the relationship between data and outcomes has changed, so models can be refreshed and remain relevant.
Compared with emerging AgentOps practices, MLOps is more established. Machine learning can also require less compute than generative or agentic approaches because it is not dependent on token-based processing.
Choose the right technology for the problem
The goal is to solve a business problem, not to force the latest technology into a use case. Starting with the tool rather than the need can add complexity without creating lasting value.
A stronger approach is to educate stakeholders on the available options, identify and prioritise the use cases, then select the most appropriate solution. That may be agentic AI, a workflow, generative AI, machine learning, a dashboard or non-AI automation. In some cases, the answer will combine several of these approaches in one solution.
The full palette matters. It includes agentic and generative AI, other generative techniques, machine learning, dashboards, website analytics and established automation. Many organisations are still developing their data maturity, so proven techniques can unlock significant value even when they are no longer considered new.
