The Next Industrial Shift: AI Agents at Work

Author: Denis Avetisyan

A new review examines how autonomous AI systems are poised to reshape engineering and manufacturing, but significant hurdles remain before widespread adoption.

The study reveals that the integration of agentic AI within engineering and manufacturing processes isn’t simply a matter of technological implementation, but rather a systemic shift-a reshaping of established workflows that either facilitates graceful adaptation or accelerates inevitable decline.

This paper details industry perspectives on the utility, challenges, and opportunities surrounding agentic AI implementation in complex engineering and manufacturing workflows.

Despite increasing investment in artificial intelligence, realizing its full potential in complex engineering and manufacturing workflows remains a significant challenge. This study, ‘Agentic AI in Engineering and Manufacturing: Industry Perspectives on Utility, Adoption, Challenges, and Opportunities’, explores current adoption trends and barriers through interviews with key stakeholders, revealing that while AI excels at automating repetitive tasks, broader implementation is hampered by data silos, verification needs, and organizational preparedness. The findings suggest a staged progression of AI utility, contingent on maturing infrastructure and trust, rather than purely technological limitations. Will overcoming these hurdles unlock a new era of intelligent automation and fundamentally reshape the future of product development and manufacturing processes?

The Erosion of Established Systems

Established engineering workflows, despite their longevity, increasingly grapple with inherent complexities and the accelerating pace of technological advancements. Traditional processes, often built upon decades-old methodologies, struggle to adapt to the demands of modern projects characterized by intricate designs, stringent regulations, and compressed timelines. This rigidity manifests as protracted development cycles, escalating costs, and a diminished capacity for innovation. The sheer volume of data generated throughout the engineering lifecycle, coupled with siloed departments and disparate software systems, exacerbates these challenges, hindering seamless information flow and collaborative problem-solving. Consequently, organizations find themselves burdened by processes that, while familiar, impede agility and responsiveness in a rapidly evolving technological landscape.

The modern engineering process often suffers from a critical weakness: data silos. Information vital to product development, testing, and maintenance frequently resides in isolated departmental databases, preventing a holistic understanding of performance and potential issues. This fragmentation necessitates laborious manual consolidation, introducing delays and the risk of human error, and ultimately impeding effective analysis. Consequently, decision-making becomes reactive rather than proactive, hindering innovation and increasing the likelihood of costly mistakes. A unified data strategy, enabling seamless information flow across all engineering functions, is therefore crucial for optimizing workflows and fostering truly informed strategic choices.

The increasing prevalence of automated workflows in engineering, while promising efficiency gains, frequently neglects the nuanced contributions of human expertise. These systems, designed to optimize predictable processes, often struggle with the ambiguity and unforeseen challenges inherent in complex projects. Critical insights – derived from years of practical experience, intuitive problem-solving, and cross-disciplinary communication – are difficult to codify into algorithms. Consequently, a sole reliance on automation can lead to suboptimal solutions, missed opportunities for innovation, and a diminished capacity to adapt to rapidly evolving circumstances, highlighting the ongoing need for synergistic human-machine collaboration within engineering disciplines.

Our team utilized a developed interview guide, as detailed in the table, to ensure consistent and focused data collection.

Agentic Systems: A Paradigm Shift in Problem Solving

Agentic AI signifies a progression from traditional AI systems by incorporating autonomous goal pursuit and action planning capabilities. Rather than solely responding to specific prompts or executing pre-defined tasks, agentic systems are designed to independently define sub-goals, strategize, and execute a sequence of actions to achieve a broader objective. This functionality is enabled through the integration of planning algorithms, reasoning engines, and iterative refinement processes, allowing the AI to adapt to unforeseen circumstances and optimize its approach over time. The core distinction lies in the shift from reactive execution to proactive problem-solving, enabling applications requiring sustained effort and adaptability beyond the scope of single-turn interactions.

Agentic AI’s ability to move beyond simple task execution facilitates a dynamic approach to engineering challenges by enabling iterative refinement of solutions. Traditional AI often requires explicitly programmed responses to predetermined inputs; agentic systems, however, can decompose complex problems into sub-goals, independently assess progress, and modify strategies based on observed outcomes. This adaptability is crucial in engineering contexts where unforeseen variables and incomplete information are common. By autonomously formulating hypotheses, executing tests – potentially through simulations or real-world interactions – and analyzing results, agentic AI can accelerate problem-solving cycles and optimize designs without constant human intervention. This capability extends to areas like automated design exploration, robotic process optimization, and the development of self-correcting systems.

Agentic AI systems differentiate themselves from conventional AI copilots by operating with a high degree of autonomy. While copilots necessitate continuous human input to define each step in a process, agentic systems are designed to independently formulate and execute plans to achieve specified goals. This capability involves internal state management, tool utilization, and iterative refinement of strategies without requiring constant external direction. Consequently, agentic AI can address complex scenarios characterized by uncertainty or evolving conditions, adapting its approach based on observed outcomes and pre-defined objectives, whereas copilots remain reliant on immediate human guidance for each action taken.

Verification Frameworks: Ensuring System Reliability

AI Verification Frameworks are essential for evaluating the performance of agentic AI systems, particularly those involving spatial reasoning tasks. These frameworks utilize a suite of tests and simulations designed to assess an agent’s ability to accurately perceive, interpret, and navigate within its environment. The complexity of spatial reasoning-requiring understanding of distance, direction, and object relationships-introduces significant potential for error; therefore, verification processes must thoroughly examine these capabilities. Current frameworks employ metrics such as path optimality, collision avoidance rates, and object localization accuracy to quantify performance. Rigorous testing within these frameworks helps identify and mitigate potential failures before deployment, ensuring system reliability and safety in real-world applications.

Current AI verification frameworks are significantly challenged by data fragmentation, stemming from disparate data sources, inconsistent formats, and a lack of standardized data schemas. This fragmentation hinders the ability to perform comprehensive validation of agentic AI, particularly in spatial reasoning tasks where accurate and complete datasets are essential. Effective mitigation requires comprehensive data integration strategies, encompassing data cleaning, transformation, and the implementation of unified data models. These strategies must prioritize data provenance tracking and version control to ensure reproducibility and accountability in the verification process, ultimately enabling a holistic assessment of AI system reliability.

Human-in-the-Loop (HITL) frameworks are implemented to mitigate risks associated with agentic AI systems by providing a mechanism for real-time monitoring and intervention. These frameworks route AI-generated outputs or proposed actions to human engineers before execution, allowing for validation and correction of potential errors or inconsistencies. The intervention process can range from simple approval/rejection of outputs to detailed modification of the AI’s reasoning or proposed actions. HITL systems are particularly critical in scenarios demanding high accuracy or safety, and they facilitate continuous improvement of the AI model through the incorporation of human feedback into the training data. Data collected from human interventions is used to refine the AI’s algorithms, reduce future errors, and enhance overall system reliability.

U.S. manufacturing employment, as reported by the Bureau of Labor Statistics, demonstrates the number of people employed in the manufacturing sector over time.

The Resurgence of U.S. Manufacturing: A New Trajectory

A significant shift is underway in U.S. manufacturing, as companies increasingly bring production back to domestic shores. This reshoring trend isn’t simply about reversing decades of outsourcing; it represents a strategic response to vulnerabilities exposed by recent global disruptions. Businesses are prioritizing greater control over their supply chains, seeking to minimize risks associated with geopolitical instability, transportation bottlenecks, and unpredictable international events. This desire for resilience is coupled with a growing recognition of the benefits of proximity to key markets and a skilled workforce, fostering innovation and faster response times. The result is a re-evaluation of long-held assumptions about cost versus control, with a growing number of manufacturers willing to invest in domestic production to ensure a more stable and dependable future.

The resurgence of U.S. manufacturing, often termed reshoring, isn’t simply a shift in location – it’s demonstrably linked to job creation within the nation. Recent analysis indicates a strong correlation between companies bringing production back to the United States and a rise in manufacturing employment figures, signaling a positive economic ripple effect. This isn’t limited to traditional factory roles; the revitalized sector increasingly demands skilled positions in areas like advanced robotics, data analytics, and supply chain management. Consequently, reshoring initiatives are fostering new economic opportunities across a broader spectrum of the workforce, potentially addressing long-standing skills gaps and bolstering regional economies previously impacted by offshoring trends.

The resurgence of U.S. manufacturing isn’t simply a matter of relocating facilities; it demands a fundamentally different operational approach, and agentic artificial intelligence is poised to deliver it. These advanced AI systems, capable of autonomous decision-making within defined parameters, promise to optimize complex manufacturing processes, enhance supply chain resilience, and address critical labor shortages. However, successful implementation hinges on robust verification protocols and continued human oversight to ensure quality control and adaptability. The current momentum-with 45% of global executives actively planning reshoring initiatives and nearly 29% already underway-underscores the urgency and potential of this technological integration, suggesting a future where AI isn’t merely automating tasks, but actively enabling a more agile and competitive domestic manufacturing landscape.

Bridging the AI Literacy Gap: A Path Forward

A critical impediment to realizing the benefits of agentic AI lies in the pervasive AI literacy gap within many organizations. This isn’t simply a lack of technical expertise, but a broader disconnect between the potential of these advanced systems and the understanding needed to effectively deploy and manage them. Without a workforce capable of interpreting AI outputs, validating its decisions, and integrating it into existing workflows, even the most sophisticated agentic AI tools remain underutilized or, worse, lead to flawed implementations. This gap extends beyond data scientists and engineers, impacting roles across the organization that require interaction with and oversight of AI-driven processes – hindering innovation and limiting the return on substantial investments in this transformative technology.

Successful digital transformation within manufacturing hinges not merely on adopting new technologies, but on cultivating a workforce proficient in their application. Current initiatives must actively prioritize comprehensive training and educational programs specifically designed to empower engineers with the skills necessary to navigate and leverage agentic AI. This extends beyond basic software training; it requires fostering a deeper understanding of AI principles, data analytics, and the ethical considerations surrounding autonomous systems. By equipping engineers with these capabilities, manufacturers can effectively integrate AI into existing workflows, optimize processes, and unlock its full potential for innovation, ultimately mitigating the projected skills gap and ensuring continued competitiveness in a rapidly evolving landscape.

The successful integration of agentic AI into U.S. manufacturing hinges on a workforce equipped to leverage its capabilities, and current data reveals a pressing need for enhanced skills. With 82% of manufacturers identifying artificial intelligence as a key growth engine – and 44% already realizing substantial returns – the potential economic benefits are significant. However, a projected shortfall of 1.9 million manufacturing engineers by 2033 threatens to stifle this progress. Closing the AI literacy gap isn’t merely about training; it’s a strategic imperative to unlock the full transformative power of agentic AI, ensuring U.S. manufacturers remain competitive on a global scale and capitalize on the rapidly evolving landscape of intelligent automation.

The study illuminates a fundamental tension within complex systems-improvements, while initially potent, inevitably succumb to the pressures of time and entropy. This echoes Henri Poincaré’s observation: “The fact that a certain phenomenon persists does not mean that it is stable; it merely means that the forces which threaten it are not yet strong enough.” Agentic AI, poised to address data fragmentation and enhance automation in engineering and manufacturing, faces a similar dynamic. While offering immediate gains in efficiency, its long-term utility hinges on navigating the inevitable decay of data integrity, the evolution of verification needs, and the persistent challenge of organizational adaptation-forces that, given enough time, will test even the most robust systems.

What Lies Ahead?

The exploration of agentic AI in engineering and manufacturing reveals a familiar pattern: the promise of automation invariably encounters the realities of systemic inertia. Data fragmentation, highlighted within this work, is not merely a technical hurdle, but a manifestation of organizational entropy-a natural consequence of complexity. Attempts to force seamless integration often resemble attempts to halt erosion with a dam; the pressure simply shifts elsewhere, revealing new vulnerabilities.

Future research must move beyond assessing what agentic AI can do, and focus on how these systems age within existing infrastructures. Verification protocols, currently a point of contention, will inevitably become more sophisticated-and correspondingly, more fragile. The pursuit of perfect validation is a Sisyphean task; the goal isn’t absolute certainty, but a manageable rate of decay.

Ultimately, the true metric of success won’t be uptime-a fleeting phase of temporal harmony-but the resilience of these systems as they navigate inevitable degradation. The question isn’t whether agentic AI will transform these industries, but whether that transformation will be a graceful adaptation or a catastrophic failure of integration. The current work suggests the latter is far more likely, unless a focus shifts from novelty to long-term systemic health.

Original article: https://arxiv.org/pdf/2604.09633.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/