The future of data science

Data science has always been a blend of statistics, engineering and business judgement. Large language models and agents are changing the blend.

Less hand-building, more framing

Many tasks that once needed a custom model — classifying text, extracting fields from documents, summarising — can now be done with a foundation model and a good prompt. The valuable work shifts to framing: which decision, which data, what counts as good, and what happens when the system is wrong.

Evaluation becomes the core skill

When systems are built from models you did not train, the question becomes: how do you know it works? Building golden sets, defining metrics, running regression tests on every change and measuring drift is now central to the job.

Production engineering matters more

A model or agent that is not integrated, monitored and owned does not change anything. Data scientists who can work with pipelines, APIs, CI/CD and observability will have outsized impact.

Classical ML is not going away

Forecasting, pricing, risk scoring and optimisation still rely on structured data and well-understood models. These are often cheaper, faster and more explainable than LLM-based approaches for tabular problems.

Governance and security become part of the role

Data scientists increasingly need to understand access controls, privacy obligations and AI-specific risks such as prompt injection and data leakage. These are now design constraints, not someone else's problem.

What this means for teams

• Pair data scientists with data and platform engineers.
• Invest in evaluation tooling and shared golden sets.
• Train teams on responsible AI and security basics.
• Measure impact by decisions changed, not models built.

Our corporate AI training programmes include practitioner labs designed around exactly these skills.