Is machine learning being overlooked?

Machine learning (ML) has quietly underpinned many of the most dependable and commercially valuable systems in modern organizations. From fraud detection and supply chain optimization to equipment maintenance. However, as attention swings toward agentic AI and large language models (LLMs), a problematic narrative is emerging that ML belongs to a previous era, overshadowed by autonomous technologies that claim to offer human-like reasoning and adaptive decision-making.

Business leaders are right to be excited about specific agentic AI use cases. The interface is intuitive, and the potential appears transformative. Research from IBM suggests that over a third (38%) of the companies they surveyed are actively implementing generative AI. But beneath the enthusiasm lies a strategic risk. Businesses may unintentionally undervalue the very ML foundations that make advanced AI systems viable, safe, and economically sustainable.

So, is machine learning being overlooked? Increasingly, the answer appears to be yes, and industry experts are growing more vocal in warning leaders not to let hype seduce them into short-term thinking.

The hype cycle is reshaping perceptions

The early appeal of agentic AI has changed the tone of board-level conversations. Large language models, wrapped in conversational interfaces, give leaders a feeling of immediate usefulness that ML systems can’t quite manage to deliver.

Manish Jethwa, CTO at Ordnance Survey, explains that this shift has created a problematic illusion. Executives now feel that AI is, “accessible” through everyday tasks like summarizing documents or drafting content, which makes some believe that structured data pipelines and ML fundamentals are no longer necessary. As he tells ITPro, agentic AI gives organizations a more intuitive interface, but it doesn’t magically add new predictive or optimization capabilities.

Others describe a deeper category error now emerging at leadership level. Eleanor Watson, an AI ethics engineer and faculty member at Singularity University, emphasized that users of these technologies increasingly equate AI only with agents and LLMs. This, she says, is misleading, because these systems still depend on classic ML principles: “Agentic systems aren’t magic – they are generative models wrapped in scaffolding.”

That lack of understanding has consequences. Overestimating the autonomy and stability of agentic systems can lead companies to underinvest in the data engineering, feature development, and ML pipelines that actually deliver measurable value. As Rafael Artacho, product director at Unit4 explains, the boom in agentic systems creates a “Swiss army knife” perception that encourages enterprises to assume LLMs can solve every problem. In practice, limitations quickly emerge when organizations try to build real, controlled solutions.

A similar concern is echoed in academia. Giannis Haralabopoulos, assistant professor in Machine Learning at Henley Business School, tells ITPro that companies often see ML as a niche data analytics tool, while AI – particularly LLMs – feels “state-of-the-art.” This perception gap, he says, is widening even as ML continues to excel, particularly when precision and efficiency matter most. In other words, ML hasn’t become less important; it has simply become less visible.

Where ML still outperforms agentic systems

If organizations follow the hype rather than the data, many risk putting investment into the wrong layer of their AI stack. ML continues to outperform agentic systems in three major domains: precision, speed, and cost-efficient scale.

High-velocity, low-latency predictions

Whether it’s fraud detection or credit scoring, businesses depend on decisions that occur in milliseconds. Autonomous agents, with their recursive reasoning loops, cannot operate at this pace or cost.

Take fintech, for example. Michele Tucci, chief strategy officer and co-founder of Credolab, says that a random forest model can evaluate whether a transaction is fraudulent at a fraction of a cent per million inferences. An agentic system, he emphasizes, might cost dollars per task, which is orders of magnitude more expensive.

The trade-off is clear: where latency and precision are non-negotiable, ML remains the only viable option.

Structured optimization and forecasting

Manufacturing and supply chain optimization rely on mathematical models with predictable behavior. Eleanor Watson, IEEE member, AI ethics engineer, and AI faculty at Singularity University, notes that these problems often have well-defined objectives and stable parameters. These are perfect conditions for ML, but ill-suited to LLM-based agents that might “reward hack” and meet goals in ways that violate business intent.

Similarly, Artacho points out that in ERP environments, precision and repeatability matter more than simulated reasoning. Financial anomaly detection, incorrect postings, and compliance-driven workflows require the determinism ML provides.

Domain-specific and non-text problems

Many business challenges include understanding sensor data operational forecasting. Haralabopoulos notes that ML still outperforms agentic systems in number-intensive prediction tasks and even certain natural language tasks like multilabel text classification.

Alex Kugell, CTO at Trio, also stresses that leaders often forget where their value really comes from. He says fraud models, demand forecasting engines, and operational optimization tools, quietly keep organizations running. And when it comes to producing precise answers based on proprietary data, ML remains indispensable.

Value away from the hype

Are companies being distracted from ML essentials?

Almost every expert ITPro spoke to expressed concern that organizations are drifting toward an “LLM-first” mindset, ignoring the foundational work required for reliable AI.

Tucci warns that companies are, “in love with the roof of the house while neglecting the foundation”. The wow factor of autonomous agents, he explained, often overshadows investments in data quality and feature engineering. The result? Agents that hallucinate or misinterpret data, leading to unreliable – and sometimes unsafe – behavior.

Watson put it even more starkly: the “wow factor”, she says, encourages deployment before readiness, creating systems that feel exciting but are prone to catastrophic errors. Because foundation models learn generic features, companies also risk assuming that domain-specific engineering is no longer needed, an assumption that she argues can lead to critical blind spots.

Haralabopoulos sees a longer-standing issue at play. Businesses, he explains, “have traditionally allowed any wow factor to distract from meaningful processes”. While cloud providers now heavily market agent capabilities, the underlying ML groundwork still relies on human expertise; and that expertise is too often left out.

Interestingly, not all experts interpret this shift as neglect. Nadine Kroher, chief scientific officer at Passion Labs, describes the ML lifecycle as having fundamentally changed. With many systems built on top of large foundation models, the classic pipeline has evolved into prompt engineering and managing dependencies on external model providers. The problem, she suggests, is not distraction but immaturity: best practices simply haven’t caught up with the new paradigm.

Even so, the message is consistent: ignoring ML fundamentals undermines the reliability of anything built on top of it, including AI.

Rethinking how ML and agentic AI work together

It’s clear that ML and agentic AI excel at different tasks and that businesses must architect systems that combine their strengths, rather than forcing one to replace the other. Artacho explains that ML brings accuracy and stability, while LLMs add context and orchestration. In his view, the strongest approach is to let ML produce calibrated signals, risk scores, anomalies and forecasts, while LLMs interpret those signals within a broader business context and propose actions.

Watson offers a sharp reminder: “It is architectural malpractice to ask an LLM to perform fraud scoring, just as it is to ask a regression model to write a customer email.” In other words, each tool must be used where it excels.

However, the depth of integration between LLMs and ML is still a challenge. Haralabopoulos remains skeptical that LLM-based reasoning fundamentally strengthens most ML pipelines. However, he acknowledges that LLMs can interpret results or improve user interaction, even if the core predictive value sits firmly within ML.

Despite different perspectives, one message is universal: ML remains the backbone of reliable AI. Agentic systems expand what is possible, but they depend entirely on this foundation.

Machine learning has never been more vital to business success. ML powers accurate decision-making and strengthens operational resilience. Agentic AI may transform how people interact with technology but without a solid ML foundation, those systems will falter.

The companies that lead in the next decade ahead won’t be the ones chasing the most dazzling agentic demos. Instead, they’ll be the ones investing in high-quality data pipelines, well-governed ML models, and hybrid architectures that combine precision with intelligent orchestration.