As bottom-line revenue from the use of LLMs continues to evade most companies, agentic AI with its purpose-driven autonomous capabilities may seem like the magic bullet for ROI.

Not so fast.  

It’s true that agentic AI is on an accelerated growth path, with Capgemini estimating that the tech, …could generate up to $450 billion in economic value,” by 2028.¹ But some of the same struggles plaguing enterprises trying to eke revenue from their generative AI (GenAI) investments – like sprawl, governance, reliability, and technical woes like drift – threaten to disrupt or even sabotage agent rollouts, as well.

Consider model drift, which occurs when the data and/or the relationships between input and output variables in a model change over time. The challenge is inherent with modeling because it stems from the assumptions that must be made during the training period. Those assumptions, the characteristics of the input data, naturally change within the lifespan of the model because fresh data is continually introduced.

A similar phenomenon occurs with AI agents, which sit atop LLMs, called Emergent behavior. When LLMs grow to large and complex, or when systems of agents grow too complex, agents can deviate from their original purpose and begin taking unpredictable actions automatically.

If a company fails to monitor and adjust for these organic and unpredictable changes, the model or agent will begin to slowly “drift” from its original parameters and begin generating inaccurate outcomes. And that results in everything from a degradation of model performance to faulty decision-making – all of which can take place without the company ever knowing it.

The challenge is only amplified in the industrial AI realm, where mission-critical systems in energy, transportation, and manufacturing, demand reliable, transparent, and observable AI. An incorrect action by an autonomous agent in industry can lead to catastrophic consequences from equipment damage to outages, to personal injury.

All of this is driving a serious lack of trust in this nascent corner of AI. Indeed, a recent McKinsey study ² noted that, “trust in fully autonomous AI agents is declining, dropping from 43% to 27% in one year. Ethical concerns, lack of transparency, and limited understanding of agentic capabilities are key barriers.”

What’s needed is an agent that organizations can trust. But how?

A time-tested approach to trust

Hitachi has been developing and delivering industrial AI solutions across digital engineering, managed services, software, data infrastructure, and more for decades. When the technical challenges surrounding agents began surfacing with customers, the company applied a methodical approach: combining reliably built agents with a secure and robust management system.

And it all started with the launch several years ago of Hitachi Digital Services’ Hitachi Application Reliability Centers (HARC) offering, a managed service platform designed to modernize and optimize cloud-based workloads.

This versatile platform quickly evolved to include new features and services, as the cloud landscape evolved. For example, earlier this year, the company added to HARC a library of AI accelerators for a wide range of industry-specific disciplines to help industrial companies jump-start their AI work.

And just recently, it expanded the platform further with a two-pronged solution to the agent problem. The brand-new HARC Agents is a blend of technologies, frameworks, and hands-on services designed to help organizations effectively deploy standardized, enterprise-class agent solutions. At its heart is an Agent Library of more than 200 agents across six key domains, and an Agent Management System with a single dashboard that centralizes control for all agentic AI platforms across an organization.

“People get very dependent on AI,” said Prem Balasubramanian, chief technical officer and head of AI at Hitachi Digital Services. “Over time, they develop greater trust in AI tools, relying on them more extensively, even for critical business operations. However, the challenge arises when these tools start to drift and Emergent behavior kicks in, silently. How will they measure this drift? How will they detect it? This is precisely where our Agent Management System comes into play.”

For its part, the HARC Agents library includes agents to help diagnose faults in machinery and vehicles, to perform quality inspections at manufacturing facilities, and to assist with financial operations, among many others. One agent even allows users to remotely control drones through conversational voice commands. But even more importantly, Balasubramanian says, the platform will help ensure those agents stay reliable and secure over the long run.

That’s because these agents and the management system join two existing offerings within the HARC platform: the R202.ai framework for defining the development and deployment of scalable, enterprise-grade AI workloads; and HARC for AI, which helps organizations operationalize and optimize AI systems.

The power of observation

There’s more to trust than management, however, says Balasubramanian. Especially in the industrial sector.

“Agents have to be reliable and responsible,” he says. “In healthcare, you can’t have similar answers. You must have the same answer every time. And hallucinations and Emergent behaviors can’t be tolerated. Agents can’t just start doing whatever they want. They must be observable, both from a cost standpoint, as well as an explainability and auditability standpoint. If an agent makes a decision or gives you a recommendation, you should be able to see why it decided this or recommended that, especially within regulated industries.”

These aren’t mere ideas. They’re baked into the methodology of the company’s R2O2.ai, which is shorthand for Responsible, Reliable, Observable, and Optimal AI.

How trustworthy agents lead to faster production

One of the related byproducts of building responsibility, reliability, and observability into such a methodical approach to agents and AI, is faster time to production. Once an organization can trust that the underlying AI and agent development is sound, they can move more confidently forward, especially through the prototype-to-production gauntlet.  

“People are realizing that prototyping is relatively straight forward,” Balasubramanian says. “However, moving to production is a different challenge, particularly for enterprises and industrial organizations. While technologists can develop the agents, the business must manage issues like anomalies and Emergent behaviors. In fact, deploying agents and establishing guardrails can constitute 70% of the effort.”

All that changes with a responsible, reliable approach. Between the HARC Agents library and the Agent Management System, the company aims to help organizations design, build, deploy, and leverage agentic AI systems in 30% less time than typically required.

Balasubramanian emphasizes the critical question organizations must now ask: Are you truly maximizing your return on investment with your current AI spending, or could you achieve greater efficiency and value by investing in agentic AI for the same workflows?

“I want all my workflows to be agented – that’s the vision,” Balasubramanian says. “With every agentic workflow, you know the price and that it’s giving me my ROI. That’s where our management system comes in. That’s where R2O2.ai comes in: optimal AI for every workflow.”

In the realm of industrial AI, moving from pilot to production, diligently monitoring performance and with a clear view of ROI is critical for mission critical systems across industries – especially in the new age of agentics.

To learn more about Hitachi Digital Services’ AI approach and HARC Agents, read: https://www.hitachids.com/service/enterprise-ai/.

And for more information about Hitachi’s industrial AI work, visit www.hitachidigital.com/ai-resource-center/.

# # #

Prem Balasubramanian is Chief Technology Officer, Hitachi Digital Services, and a Hitachi Ltd. Global AI Ambassador.

Hitachi Digital Services, a wholly owned subsidiary of Hitachi, Ltd., is a global systems integrator powering mission-critical platforms with people and technology. It helps enterprises build, integrate, and run physical and digital systems with tailored solutions in cloud, data, IoT, and ERP modernization, underpinned by advanced AI.

¹Capgemini: Rise of Agentic AI: How trust is the key to human-AI collaboration https://www.capgemini.com/insights/research-library/ai-agents/
²McKinsey: QuantumBlack: https://www.mckinsey.com/capabilities/quantumblack/our-insights/seizing-the-agentic-ai-advantage