Every tech conference this year tells the same story. Keynote speakers unveil dazzling AI agents that book flights, write code, and orchestrate entire business workflows with minimal human intervention. Venture capital floods into agentic AI startups. CEOs pepper earnings calls with references to autonomous systems that will reshape their companies from the inside out.
Meanwhile, the technology actually generating returns for most enterprises right now is decidedly less glamorous. It’s the logistic regression model flagging fraudulent credit card transactions. The recommendation engine nudging e-commerce customers toward their next purchase. The demand forecasting algorithm that keeps warehouse shelves stocked at optimal levels. Classical machine learning — the kind that’s been around for years, sometimes decades — remains the financial backbone of enterprise AI.
This isn’t a contrarian hot take. It’s what the data shows.
As TechRadar reported, the growing obsession with AI agents risks obscuring a fundamental reality: traditional machine learning models still pay the bills at most organizations. The piece, drawing on insights from Evo Pricing CEO Fabrizio Fantini, argues that enterprises chasing the latest agentic AI trend may be neglecting the proven, revenue-generating ML systems already embedded in their operations. Fantini’s core message is blunt — companies should be investing more in the machine learning infrastructure they already have, not abandoning it for technologies that remain largely experimental in production settings.
The distinction matters more than the industry’s marketing apparatus would suggest. Machine learning, in its classical form, encompasses supervised and unsupervised learning techniques — decision trees, gradient boosting, support vector machines, neural networks trained on structured data — that excel at well-defined prediction tasks. These models take historical data, identify patterns, and produce outputs that feed directly into business decisions. Pricing optimization. Churn prediction. Credit scoring. Supply chain management. The applications are mature, measurable, and, critically, profitable.
AI agents represent something different. They’re systems designed to perceive their environment, make decisions, and take actions autonomously, often chaining together multiple AI capabilities — large language models, retrieval systems, code execution — to accomplish complex, multi-step tasks. The promise is enormous. The production reality, so far, is considerably more modest.
That gap between promise and production is where enterprises keep stumbling.
A McKinsey Global Survey published earlier this year found that while 72% of organizations have adopted AI in at least one business function — the highest rate the consultancy has ever recorded — the vast majority of deployed use cases still rely on traditional ML and analytics rather than generative AI or agentic systems. The revenue impact from these conventional models continues to grow steadily, even as investment attention shifts elsewhere.
Gartner has echoed similar findings. The research firm’s 2024 analysis of enterprise AI deployments showed that predictive analytics and traditional ML models account for the majority of AI-driven business value, and that most organizations attempting to deploy AI agents are still in pilot or proof-of-concept stages. The firm warned that enterprises risk “innovation fatigue” by constantly pivoting to new AI paradigms before extracting full value from existing investments.
So why does the industry keep chasing the next shiny thing?
Part of the answer is structural. Vendor economics favor novelty. Cloud providers, software companies, and consultancies generate more revenue from selling new platforms and services than from helping clients optimize what they already have. When Microsoft, Google, and Amazon Web Services are all racing to position their AI agent frameworks as the next essential enterprise tool, the marketing pressure on CIOs and CTOs is immense. Nobody wants to tell their board they’re focused on improving a five-year-old gradient boosting model when competitors are announcing autonomous AI initiatives.
But the smartest operators know better. As Fantini told TechRadar, the real competitive advantage often lies in refining existing ML systems — better feature engineering, cleaner data pipelines, more frequent retraining cycles, tighter integration with business processes. These improvements compound over time. A 2% improvement in a demand forecasting model’s accuracy might translate to millions in reduced inventory waste across a large retail operation. That’s not hypothetical. It’s happening right now at companies that resist the temptation to reallocate their ML engineering teams to agent-building projects.
The reliability question looms large, too. Traditional ML models, when properly validated and monitored, behave predictably. They produce numerical outputs — a probability score, a price recommendation, a classification — that can be tested, audited, and explained to regulators. AI agents, by contrast, introduce layers of unpredictability. An agent that chains together multiple LLM calls, tool uses, and decision points creates a combinatorial explosion of possible behaviors. Debugging a failure in a ten-step autonomous workflow is fundamentally harder than diagnosing why a random forest model misclassified a transaction.
This isn’t an argument against AI agents. Not remotely.
The technology is genuinely advancing. OpenAI’s agent-oriented releases, Anthropic’s Claude-based tool use capabilities, Google DeepMind’s Project Mariner, and a wave of startups building vertical-specific agent platforms are pushing the boundaries of what autonomous systems can accomplish. In customer service, software development, and certain back-office functions, agents are beginning to deliver measurable value. The trajectory is clear: agentic AI will become increasingly important over the next several years.
But trajectory and current reality are different things. And conflating them leads to misallocated budgets, disappointed stakeholders, and — perhaps worst of all — neglected ML systems that quietly degrade when starved of engineering attention. Models drift. Data pipelines break. Feature stores go unmaintained. The irony is acute: companies let their proven revenue generators atrophy while pouring resources into experimental systems that may not reach production for another 18 months.
The enterprise AI market is projected to exceed $300 billion globally by 2027, according to IDC estimates. The overwhelming majority of that value, at least in the near term, will come from traditional ML applications — fraud detection, personalization, operational optimization, predictive maintenance — rather than autonomous agents. Companies that recognize this aren’t being conservative. They’re being rational.
There’s a useful analogy in the history of enterprise software. When cloud computing emerged in the late 2000s, the hype cycle was deafening. Every vendor pivoted to cloud. Every analyst predicted the imminent death of on-premises infrastructure. And yet, more than 15 years later, hybrid architectures dominate. On-prem workloads still represent a significant share of enterprise IT spending. The cloud transition happened — and continues to happen — but far more gradually and messily than the narratives suggested. Traditional ML and AI agents will likely follow a similar pattern of coexistence, with the older technology retaining its importance far longer than today’s conference keynotes imply.
Fantini’s argument, as captured by TechRadar, carries an additional nuance worth examining. He suggests that the best path to effective AI agents actually runs through strong classical ML foundations. Agents that need to make pricing decisions should be built on top of well-tuned pricing models. Agents that manage supply chains should incorporate battle-tested demand forecasting systems. The agent layer adds orchestration and autonomy; the ML layer provides the analytical intelligence. Strip away the ML foundation, and you’re left with an agent that can talk fluently but can’t actually think quantitatively.
This layered approach — classical ML as the analytical engine, agentic frameworks as the orchestration layer — is gaining traction among enterprise architects who’ve moved past the initial hype. It acknowledges that these technologies aren’t competitors but complements, with very different maturity curves and risk profiles.
The talent dimension compounds the challenge. Experienced ML engineers who understand feature engineering, model validation, and production deployment are in shorter supply than ever, partly because many have been redeployed to generative AI and agent projects. Junior engineers entering the field increasingly want to work on LLMs and agents, not on optimizing XGBoost models or building data pipelines. The result is a growing maintenance burden on the systems that actually drive revenue, staffed by shrinking teams.
Smart companies are addressing this by creating explicit career tracks and investment frameworks that value ML operations alongside AI innovation. They’re treating their existing model portfolio the way a financial firm treats its existing book of business — as an asset requiring ongoing management, not a legacy liability to be abandoned.
None of this means enterprises should ignore AI agents. The technology is real, the potential is substantial, and early movers in certain verticals will gain advantages. But the allocation of attention and resources should reflect where value is actually being created today, not where it might be created in 2027. For most organizations, that means doubling down on classical ML while running disciplined, well-scoped agent experiments on the side.
The unsexy truth of enterprise AI in 2025 is that the most profitable models in production at most companies would bore a conference audience to tears. They’re not chatting with customers or autonomously managing workflows. They’re crunching numbers in batch jobs that run at 3 a.m., producing predictions that flow into dashboards and decision systems that humans act on every morning.
And they’re making real money. Quietly. Reliably. Every single day.
That should count for something — especially when the board asks where the AI budget went.
WebProNews is an iEntry Publication