Real AI ROI: Why Traditional ML Techniques Still Win

If you're an organisation that wants to be seen implementing AI but also want to get real long term ROI from your project you need to be looking at the 'old school' techniques. I'm talking about classifiers and regression techniques. Not only do these counter a lot of the downsides of LLMs, like context limitations and hallucinations they're also significantly more mature meaning you actually know what you're getting and can rest easy when they're in production.

The Hybrid Approach

I'd choose a 'old school' use case that can be enhanced by a modern LLM doing what it does best- information analysis and interpretation. For example, deploy a classifier to classify customer transactions as fraudulent or not. Than use the information about that transaction and customer as the context of a foundation model call. The generated analysis and raw data can than be shown to an operator for further action. This way you get both the reasoning benefit from the LLM as well as the big data analysis of the classifier.

Modern Efficiency Gains

In the past, a ML project like this has involved, data engineering, model training, infrastructure and ops. Traditionally this has been expensive but today you're looking at 2x-10x upticks in efficiency in each step - think cursor for ETL job development and Sagemaker autopilot for model training + selection. Not only this but going through this process will likely bring real long term value to your organisation's data, data governance and data collection capabilities.

Looking Ahead

Looking ahead, as model capabilities and organizational maturity evolve, these workflows will shift toward direct model calls and AI-orchestrated training. We're not quite there yet, but adopting this foundation now positions you for a significant head start.