■ MODERATE RISK ■ Technology
Transformed rather than replaced, though large parts of the classical workflow are already automated. The job is migrating from training models to operating them, which is a shift in content rather than a decline in demand.
“Ironic. You build the thing that replaces everyone else. For now, you're safe — you're the arms dealer.”
Our AI replacement risk score — how we score jobs
The classical role covers the pipeline from data to production: assembling and cleaning training data, feature engineering, model selection and hyperparameter tuning, building evaluation sets, then the harder half, which is deployment, serving latency, monitoring for drift, retraining schedules, A/B testing and the on-call responsibility when a recommender starts serving nonsense. In practice a large share of the calendar goes to data plumbing and infrastructure rather than to anything resembling modelling.
Automation has genuinely eaten the middle. AutoML and automated hyperparameter search made manual tuning largely pointless years ago, feature stores standardised much of the engineering, and managed platforms handle serving, scaling and monitoring with far less bespoke code. More disruptively, foundation models replaced entire categories of bespoke model building: classification, extraction, summarisation and even some forecasting tasks that once justified a training project are now an API call plus a prompt and an evaluation harness. Coding agents also write the pipeline code that used to fill the week.
Demand nonetheless keeps rising, because putting a probabilistic system into production reliably is difficult in ways that resist generalisation. Someone must define what good looks like for a specific business problem, build the evaluation set that catches the failures that matter, decide when to fine-tune versus prompt versus retrieve, control inference cost, handle latency budgets, detect drift in live traffic, and deal with the messy politics of a model whose errors have regulatory or reputational consequences. Data quality remains a human archaeology project. Our score of 40 says the craft changes substantially by the 2030s and the headcount does not collapse, because the constraint on deploying machine learning was never the modelling code.
Automatability: our editorial assessment of current and near-term AI capability
Serious change by around 2030, weighted toward role redefinition. Bespoke model training for common tasks keeps shrinking as foundation models absorb them, while demand grows for engineers who can evaluate, serve, monitor and cost-control AI systems in production. Expect the title to fragment into applied AI engineering, evaluation and safety work, and infrastructure specialisms. Anyone whose skill set is scikit-learn plus hyperparameter search will feel it earlier than the rest.
Yes, though a different one than five years ago. Fewer roles involve training bespoke models for tasks a foundation model now handles; far more involve integrating, evaluating, serving and monitoring AI systems inside products. Total demand is rising because more companies are deploying, but the skills that get hired have shifted toward production engineering and evaluation.
It removed a chunk of the work, specifically hyperparameter search and much of model selection, which few people miss. What it did not touch is framing the problem, sourcing and cleaning data, deciding what counts as success, and running the thing reliably against live traffic. Those were always the expensive parts of a project.
They threaten one version of it. If your value was training classifiers and extractors from scratch, an API call now does much of that adequately. If your value is knowing when a general model is insufficient, building evaluations that catch its failures, controlling cost and latency, and keeping systems honest in production, the same shift is a substantial source of demand.
Evaluation methodology first, because everything else depends on knowing whether output is good. Then inference infrastructure and cost, retrieval architectures, monitoring for drift and regression, and enough domain knowledge to weight errors correctly. Classical modelling remains useful, but it is now one tool among several rather than the centre of the job.