Could agentic AI soon become as fundamental to business operations as the PC was in the 1990s? That’s the big question that many experts have been debating recently and trying to answer.
To be clear, there is much fuss about what agentic AI really means across the industry, with some even arguing that it’s fast becoming another over-hyped buzzword. Some others, like Ryan Salva, a senior director of product management at Google and former VP of product at GitHub, say they’ve come to “hate the word agents.” Salva told TechCrunch in a recent interview that he thinks “the industry overuses the term ‘agent’ to the point where it is almost nonsensical.”
Yet, on some base levels, there seems to be an industry-wide consensus on the fact that unlike generative AI — which focuses on generating texts, images, videos, and audios at scale — AI agents are designed to take action, making decisions and executing tasks with increasing levels of autonomy.
But, really, how big is the value of agentic AI to businesses across the world today? And are they ready to handle the way we do business, especially as the AI boom continues to be faced with challenges of energy consumption, learning efficiency and data reliability?
AI’s Insatiable Energy Demand
Technology hypes are always exciting, but when the dust settles, real, hard questions begin to come to the fore. The same is the case with agentic AI, which has now gained widespread attention.
The truth is that the widespread deployment of AI agents hinges on solving the growing energy crisis tied with AI. As I’ve noted in several previous articles on Forbes, the industry is already grappling with the immense power needs of GPUs, with hyperscalers even exploring nuclear energy to sustain AI development.
“The massive CapEx investments in GPUs and AI infrastructure today echo past industrial revolutions, where foundational technologies reshaped economies,” said Amit Walia, CEO of Informatica. “And while hardware is critical, energy efficiency will be a defining factor in AI adoption,” he added
AI models, particularly those requiring real-time decision-making, demand vast amounts of processing power. This means that companies that cannot optimize their AI infrastructure risk unsustainable operational costs.
Walia, however, pointed to energy-efficient AI models as a critical step in addressing the AI energy problem. “Efficient AI agents, that use less power, lower operational expenses and align with environmental goals, will make them more appealing,” he told me.
While AI’s massive energy demand is a major challenge to AI development today, several more problems lie on the horizon.
The Need For Smarter Learning Algorithms
Beyond infrastructure concerns with AI data centers and energy consumption, AI agents must have the capability to learn and adapt in ways that exceed traditional AI models.
Thankfully, reinforcement learning — which, according to Walia, allows AI agents to refine their behavior over time, using real-world and synthetic data to simulate different scenarios — is emerging as a key enabler in that regard, allowing AI agents to refine their decision-making based on a trial and error process rather than relying solely on pre-programmed outputs.
“Generative AI models mainly rely on transformers to convert natural language inputs into outputs, but AI agents require something more: the ability to learn from experience and make decisions autonomously,” noted Walia.
Srinivas Njay, CEO of Interface.ai, echoed Walia’s sentiment, noting that reinforcement learning, often shortened as RL, is indispensable for AI agents executing complex tasks.
“For agentic AI — which must execute tasks end-to-end — RL enables the AI to navigate decision trees, adapt to changing conditions, and continuously improve by learning from successes and mistakes,” Njay explained. “Instead of just generating an answer, the agent learns to act in ways that deliver tangible business outcomes.”
But while RL enables AI to refine decision-making dynamically, it is far from a silver bullet. RL has several limitations, including but not limited to high data and compute costs, lack of interpretability about why a model made a particular decision and poor transfer learning, especially as RL models trained in one environment often struggle to adapt to new situations without significant retraining.
That explains why many cutting-edge AI applications now integrate RL with supervised learning, unsupervised learning, and retrieval-augmented approaches like RAG to overcome RL’s limitations.
It’s really quite simple: As AI continues to evolve, we will need smarter and smarter algorithms, or else we risk redundancy. And if we consider that most of today’s AI models do arguably the same things, with some doing better than others on a few benchmarks, one could argue that the redundancy is already happening.
The Data Challenge
While data remains the cornerstone of AI performance, it’s also the biggest bottleneck for agentic AI. AI agents are only as good as their training data and without high-quality, domain-specific data, they cannot function effectively in industry-specific environments like healthcare, finance and customer service.
“Our recent CDO report found that 43% of businesses cite data quality, completeness and readiness as their biggest obstacle in deploying AI initiatives,” said Walia who also emphasized that “without high-quality, domain-specific data, even the most advanced AI models will fall short.”
Njay sees this challenge firsthand in financial services, where AI agents are being tested in areas such as online banking and fraud detection. “Data silos, regulatory constraints and inconsistent formats make it difficult for AI to act with confidence,” he said.
But to solve that problem, according to Njay, “the key is modernizing data infrastructure, unifying silos and ensuring real-time access to high-quality information.”
Reality Check
As fad gets separated from reality, amidst the excitement, it has now become more apparent that most businesses aren’t yet prepared to fully hand over decision-making to AI agents — especially in high-stakes scenarios involving customer relationships, financial transactions, or strategic planning.
“There is no doubt AI agents will transform business, but right now, they excel in structured, repetitive tasks,” said Walia. “Where the hype outpaces reality is in high-stakes decision-making. At the end of the day, things get created by humans. And managing those people matters. AI agents aren’t yet ready to fully manage complex client relationships or operate without human oversight.”
Njay also agrees, noting that while AI can handle workflows like dispute resolutions and loan applications, regulators and executives alike will demand human oversight for more complex processes. “The guiding principle is ‘AI for tasks, humans for judgment’,” he said. “AI agents excel at rules-based processes, but humans provide the guardrails — especially in exceptions where trust and empathy matter.”
Choose Strategy Over Hype
We are at the beginning of what could be a long transformation, similar to the rise of enterprise software and cloud computing. While AI agents will undoubtedly become more capable, businesses must first focus on the fundamentals: ensuring data readiness, improving AI literacy among staff and integrating AI in ways that drive measurable productivity gains.
“The pieces are coming together — advancements in GPU efficiency, reinforcement learning, and domain-specific data are all accelerating adoption,” noted Walia. “The businesses that master their data management and collaboration with AI agents will be best positioned to lead the way.”
To be honest, it’s not yet clear when the core ingredients for AI agents to really transform business operations will fall into place, or how long that would take. But until then, the companies that focus on strategic AI deployment rather than chasing hype will be the ones reaping the real rewards.
