You would be forgiven for thinking that something as frequently discussed as machine learning could actually be functionally ignored.
Still, here we are. Machine learning seems to have slipped from its rightfully earned pedestal. The current situation is almost incomprehensible.
Over the past two years, agent systems have gained traction and continued to draw attention to AI. These agents promised, and in some cases delivered, autonomous workflows and natural language orchestration.
Article continues below
But it doesn’t matter. Regardless of what these AI tools did or didn’t do, the spotlight remained on them, and machine learning has suffered because of it.
Co-founder and Chief Strategy Officer of Credolab.
Machine learning never really lost value. But the industry has forgotten what it actually takes to operationalize AI within a real business, and an LLM alone can’t do it.
It’s worth remembering that machine learning didn’t go anywhere. In fact, this is the core technology used to create and run LLM. In the AI hierarchy, large-scale language models are a specialized subset of deep learning, which itself is a subset of machine learning.
When you peel back the layers, what’s left is ML, which does the heavy lifting.
Executives distracted in a strange new world
Management’s perception of AI can be divided into three distinct phases. Until around 2023, AI had an element of novelty to bright-eyed executives. The first GenAI demo certainly felt like magic.
Executives watched the models write code, summarize reports, and generate images. Although impressive, its exact operational use remained somewhat unclear. This did not prevent them from adopting it.
At some point around mid-2023, executives stopped asking whether they should use it or not, and started asking how many different areas they could poke into to determine whether it was needed or not.
These decision makers have leapfrogged their company’s AI strategy from a “tool that helps in specific situations” to a “competitive checkbox for everything that pleases shareholders.” There were many large-scale pilots with vague KPIs (and a severe lack of measurable ROI).
That lasted until early 2024, then reality set in and we found ourselves at the current stage. Pragmatism is almost back. Executives are reluctantly recognizing that while AI agents are attractive, they are expensive to operate, difficult to control, and poorly suited for high-precision decision-making.
As is typical of such sobering moments, many organizations were forced to remember their fundamental values. Traditional ML models that quietly power data quality, feature engineering, and high-speed digital operations are better than ever.
Have you always been ML?
The LLM-centric hype cycle that has dominated AI for the past few years has not taken away ML’s quiet dominance in a few select areas.
Although this is not a competition, traditional ML models still significantly outperform agent systems in high-speed, low-latency, high-precision decision-making environments. Gradient boosting machines, random forests, and logistic regression do things that agent systems cannot do.
Agents cannot “think through” whether a payment transaction processed through a fintech is fraudulent without creating a ridiculous invoice. To do this cheaply, we need a random forest that can generate a probability score of just 1 cent per million inferences in milliseconds.
Behind the scenes, ML makes sense in every way. In e-commerce, ML determines recommendations, pricing, and inventory levels at scale. In cybersecurity, anomaly detection models scan millions of data points per second. This far exceeds the best capabilities of LLM-based agents.
In mobility and logistics, ML predicts demand, optimizes routing, and adjusts allocation with tight delay constraints. If you want to give AI the real backbone for operational decision-making, you’ll rely on traditional ML instead of agents.
If you ignore the fundamentals of ML, you do so at your own risk.
Many organizations today suffer from what I call the Sistine Chapel Syndrome. Everyone’s attention is focused on the dazzling ceiling, and no one pays attention to the foundations that keep everything so high. This is a problem for organizations where the marble beneath their feet is beginning to crack.
The “garbage in, garbage out” rule didn’t go away because the interface looked like a chatbot. In fact, it got worse. Agents given bad data not only return incorrect numbers, but also confidently tell plausible stories that lead the team to poor decisions.
If the underlying ML process is weak, the agent will amplify that weakness rather than compensate for it. That ceiling crumbles when companies skip data quality checks, neglect feature engineering, and fail to properly monitor and retrain their pipelines. The rise of agents should prompt organizations to double down on ML hygiene.
two is better than one
AI is not a zero-sum game between ML or agents. The best teams work together. In the words of Daniel Kahneman, ML should be fast, automatic, accurate, and take care of prediction and classification. The LLM breaks down reasoning, orchestration, and interpretation, so you need to be time-consuming but smart.
That split should not be 50/50. Numbers are not important here at all, as the right balance depends on the results. Success should be measured by measurable business impact (such as an X% increase in operational efficiency), not just how sophisticated the underlying model is.
ML is cheap, reliable, and easy to audit. It already supports most mission-critical decisions. Any agent is only as good as the ML layer below it. This split might even save them trouble, as regulators are keen on the kind of explainability they can’t glean from probability-driven token generators.
ML, again
The flashiest agents won’t send you racing past their competitors. A clean, well-managed data pipeline built on strong ML baselines can do just that. Add in robust monitoring and retraining cycles, clear interfaces between agents and predictive models, and strong cultural discipline around model governance, all of which will pay off over time.
Investing in an invisible layer is never sexy, but those of us who overlooked machine learning have learned that it’s always essential. Perhaps ML has simply been overlooked because it works so well that it has faded into the background. No matter where the industry’s attention turns next, the overwhelming part of value creation will continue to be ML’s ability to make reliable predictions at scale.
ML maintains everything. More leaders need to realize that the path to powerful agent AI is only possible through a strong ML foundation. If you want to keep your ceiling in luxurious condition, it is better to start by repairing cracks in the floor.
Check out our list of the best IT automation software.
