The financial reality is stark. Anthropic’s CEO Dario Amodei has stated publicly that frontier AI model training runs from roughly $100 million at the lower end to over a billion dollars for models currently in development, with projections reaching $10 to $100 billion within the next few years. Even smaller, domain-specific models carry significant costs: training a seven-billion-parameter model runs $50,000 to $200,000 in compute alone, and a twenty-billion-parameter model can reach $2 million before a single line of production code is written.

The assembly line that makes Tier Three work requires three distinct, highly specialised roles that cannot substitute for one another. Misunderstanding this is where most organisations stumble.

The Data Engineer is the miner and logistics architect. Before any model can learn, data must be extracted from scattered legacy databases, cleaned of errors and duplicates, formatted consistently, and transported into a centralised warehouse. Raw enterprise data is almost never ready for machine learning. It is messy, incomplete, contradictory, and riddled with encoding artifacts from decades of system migrations. Research consistently finds that 70% to 85% of AI project failures trace to data quality issues, not algorithmic shortcomings. MIT research shows 82% of machine learning projects stall specifically because the data is not ready. The old computer science aphorism holds with brutal precision: feed bad data into a sophisticated model and you get a confidently wrong sophisticated model.

The Data Scientist is the blueprint architect. Once clean data arrives in the warehouse, someone must determine what is actually possible to predict, classify, or automate. Data Scientists design controlled experiments, test mathematical algorithms, and build the theoretical framework for how the model should learn. They are statisticians, researchers, and mathematicians, and their job is to ensure the model is solving the right problem with the right method, not simply memorising noise in the training set.

The Machine Learning Engineer is the factory manager. A Data Scientist’s prototype might perform beautifully on a laptop with a curated test dataset. The ML Engineer takes that prototype and turns it into production-grade software: managing the training runs on expensive cloud GPU clusters, optimising the model for speed and memory efficiency, deploying it into infrastructure that serves thousands of concurrent requests without crashing, and building the monitoring systems that detect when the model’s performance degrades over time. Research shows 91% of ML models experience temporal degradation, meaning the work does not end at deployment, it shifts to continuous retraining and validation.

The honest question every executive must ask before entering Tier Three is not “Can we do this?” but “Must we do this?” You only build a proprietary model when the AI is your product and you possess genuinely unique data that no commercially available model can replicate. If your competitive advantage lies in how you apply intelligence rather than in the intelligence itself, Tier Two is almost certainly the correct answer, and you will reach the market years faster.