Heterogeneous integration is a key enabler of today’s AI innovations. By bringing together multiple chips with different functionalities, a.k.a., chiplets, AI devices have been able to achieve tremendous performance gains. However, the heterogeneous integration of advanced packages has its own set of process control obstacles that must be addressed, including new interconnect challenges involving redistribution layers (RDL) and bond pads.
Recently, Onto Innovation and Samsung Electronics Co., Ltd., teamed up to explore how picosecond ultrasonic technology could be used to measure the metal thickness of RDL and bond pads in high performance AI packages. In this blog, the second in our series on the advanced packaging applications of picosecond ultrasonic technology, we will show how this technology can be used to measure metal films during RDL and bond pad processes.
But first, a word about picosecond ultrasonic technology, a widely adopted non-contact, non-destructive acoustic technique that can be used to measure film thickness.
Measuring films
Picosecond ultrasonic technology measures film thickness by tracking the round-trip travel time of ultrasonic waves generated and detected using an ultrafast laser pump probe technique. A short laser pulse (pump) creates an acoustic wave that travels through the film, reflects at material interfaces, and returns to the surface. A second laser pulse (probe) detects the returning wave.
Two detection methods can be used to determine film thickness or properties:
REF mode senses changes in surface reflectivity caused by the returning wave.
PSD mode detects surface deformation by measuring shifts in the reflected probe beam.
By measuring the time it takes for the wave to return and knowing the speed of sound in the material, the film thickness can be accurately determined to sub-angstrom levels.
This level of layer-specific metrology, precision, and measurement repeatability is increasingly critical as AI-driven packaging pushes the limits of interconnect density and uniformity.
Accuracy and repeatability
For the purpose of our exploration, we conducted a test to confirm the accuracy of picosecond ultrasonic technology when measuring the films typically used in advanced packaging. These metals include Au, Ni, physical vapor deposition (PVD) seed Cu, and RDL Cu (EP). For each film, we used picosecond ultrasonic technology to measure wafers of varying thicknesses. Then we cut the wafers for cross-section analysis and estimated the correlation with the picosecond ultrasonic results for the four films (Figure 1). In this scenario, the correlation factor R2 was higher than 0.99 for all four cases, with the slope close to one, demonstrating the accuracy of picosecond ultrasonic measurements.
This level of correlation is not only impressive, it is essential. Competing technologies such as four-point probe (4PP) or contact profilometry often fall short in multilayer structures or non-planar surfaces, where mechanical contact can distort results or damage delicate features.
Fig. 1: Correlations between picosecond ultrasonic measurements and cross-section analysis for Au, Ni, seed Cu (PVD), and RDL Cu (EP). The excellent correlation factors demonstrate the accuracy of picosecond ultrasonic technology.
Following this, we measured product wafers in various interconnect processes with picosecond ultrasonic technology, including seed Cu/Ti measured in REF mode (Figure 2) and RDL in PSD mode (Figure 3). RDL thickness can be measured both in pre- and post-seed Cu removal.
Fig. 2: Measurement signal of seed Cu/Ti in REF mode. Delay time for seed Cu and Ti are indicated by the red arrows.
Fig. 3: RDL Cu signal after the seed Cu etch process. The red arrow shows the round-trip time of an acoustic wave within RDL Cu film.
The horizontal axis in Figures 2 and 3 represents the time delay of the probe pulse with respect to the pump, while the vertical axis represents the change of reflectivity (ΔR/R) caused by the travelling acoustic wave. The sharp change of reflectivity in the signal, as demonstrated in Figures 2 and 3, is mostly due to the acoustic wave reflected from the film interface returning to the surface. In addition, the position of the peak and trough is shown with red arrows. These arrows are directly related to the thickness of the films, seed Cu, barrier Ti, and EP Cu. From the position of the peak and trough, the thickness of each film can be calculated. For seed Cu and barrier Ti, the repeatability of each layer is 0.3% or less of the thickness for all measurements. This demonstrates the capability of picosecond ultrasonic technology to meet 10% gage repeatability and reproducibility requirements.
For RDL Cu, the sharp change of reflectivity near 2,200 picoseconds (ps) corresponds to the round-trip time of the acoustic wave within the RDL Cu film; Cu thickness can be calculated from the trough position. The sharpness of the trough, along with thickness, indicates the trough position can be calculated with good repeatability. In fact, the repeatability of RDL Cu measurements for each point is less than 0.1% of Cu thickness, once again exceeding the 10% gage repeatability and reproducibility requirements.
Such precision is a necessary technical achievement. As AI applications demand tighter control over signal integrity and power efficiency, the margin for error in interconnect thickness shrinks dramatically. Legacy tools simply cannot keep up.
Measuring bond pads with dimple structures
We also used picosecond ultrasonic technology to measure a bond pad with a dimple structure. The film stack is composed of Au/Ni/Cu, with Au being the top film. Although the height of the center region of the pad is lower than the surrounding region by a few microns, we successfully measured individual layer thicknesses by measuring a few sites in the outer ring area and selectively choosing ones with good signal-to-noise ratios. This is possible because the focused spot size of the picosecond ultrasonic beam is 8×10µm2, small enough for the direct measurement on the outer ring of the pad.
This is another area where contact-based methods struggle. The ability to selectively target small, non-planar regions without physical interference is a key differentiator of picosecond ultrasonic technology.
Fig. 4: An example of an REF mode signal from the bond pad with a dimple structure for Au (a), Ni and Cu (b).
In Figure 4 a-b, the red arrows indicate the reflectivity changes caused by the acoustic waves returning from the interface to the surface. With these peak positions, we were able to calculate each layer’s thickness with good accuracy and repeatability. The repeatability of Au, Ni and Cu films for each measurement was less than 0.2%, 0.05% and 0.05%, respectively. As such, all three film measurements outperformed the requirement of 10% gage repeatability and reproducibility.
It should be noted that Au film is much thinner than the other two films. As such, there is a significantly higher repeatability for Au films compared with the other films.
Conclusion
The AI era is upon us, and it would not be possible without advanced packaging innovations. However, the complexity of today’s advanced packaging is worlds away from traditional packaging. Today’s back-end process involves a variety of technologies and requires new methods of process control. In addition, controlling metal thickness and within wafer uniformity in these processes is critical to meeting the requirements for signal integrity in advanced packaging.
Unfortunately, some fabs still rely on legacy metrology tools like 4PP or contact profilometry—technologies that were never designed for the complexity of modern AI packages. These tools often introduce mechanical stress, lack the resolution to resolve thin or buried layers, and cannot reliably measure non-planar or dimpled structures.
As we have demonstrated, picosecond ultrasonic technology is an ideally suited interconnect metrology solution for both RDL and bond pads. This technology offers excellent accuracy and gage capability for the control of interconnect processes in advanced packaging.
As back-end processes demand the same level of precision, uniformity, and control traditionally associated with front-end requirements, picosecond ultrasonic technology can play a major role in advanced packaging metrology, delivering the non-contact, high-resolution, and repeatable measurements that AI applications demand.
Acknowledgments
We would like to thank Dae-Seo Park, Sanghyun Bae, Junghwan Kim, and Hwanpil Park of Samsung Electronics Co., Ltd., and Kwansoon Park, G. Andrew Antonelli, Robin Mair, Johnny Dai, Manjusha Mehendale and Priya Mukundhan of Onto Innovation for their contributions to this article.
MACo’s 2025 Summer Conference Solutions Showcase features thought leaders and industry experts presenting resources and best practices to assist local governments. Check out this session hosted by Kaiser Permanente!
About Kaiser Permanente: Founded in 1945, Kaiser Permanente is recognized as one of America’s leading healthcare providers and not-for-profit health plans. They currently serve more than 11.3 million members in eight states and the District of Columbia. Care for members and patients is focused on their total health and guided by their personal physicians, specialists, and team of caregivers. Their world-class medical teams are supported by industry-leading technology advances and tools for health promotion, disease prevention, care delivery, and chronic disease management. Kaiser Permanente (KP) exists to provide high-quality, affordable healthcare services and to improve the health of its members and the communities it serves. They are trusted partners in total health, collaborating with people to help them thrive and creating communities that are among the healthiest in the nation.
SOLUTIONS SHOWCASE SESSION:
Title: Reimagining Care: How AI and Technology Are Powering the Future of Healthcare
Description: This session explores how a large integrated healthcare organization uses technology and AI to improve care delivery and clinical outcomes. Speakers will discuss responsible AI—focusing on ethical, transparent, and explainable tools—and why eliminating bias is essential to building trust and delivering equitable care. Our goal is to strengthen patient connections, surface relevant insights at the point of care, and support better care decisions by equipping clinicians with smart, intuitive tools. By integrating advanced technology into clinical practice, Kaiser Permanente is making care more seamless, efficient, and high-quality. Join this session to learn how these innovations are helping Kaiser Permanente enhance the patient experience and transform how healthcare is delivered.
MACo’s Summer Conference, “Funding the Future: The Evolving Role of Local Government,” will be held at the Roland Powell Convention Center in Ocean City, MD, on August 13-16, 2025.
AI is enhancing productivity at Microsoft – and could be threatening jobs after all
The company says it saved $500 million by using AI in call centers
Microsoft has laid off thousands of workers in 2025 alone
Microsoft has declared that artificial intelligence is now saving the company money across sales, customer services and software engineering.
Reports have claimed that in a recent company meeting, Microsoft’s Chief Commercial Officer Judson Althoff revealed the company has saved over $500 million in its call centers alone, thanks to the implementation of artificial intelligence, while simultaneously improving employee and customer satisfaction.
AI’s direct effects on the workforce remain uncertain, but Microsoft has laid off thousands of workers recently since overhiring during the pandemic, and it seems AI-induced efficiency gains have only worsened the effects.
Microsoft is making huge savings thanks to AI
Artificial intelligence is now handling Microsoft interactions with smaller customers, generating tens of millions in revenue with reduced human input.
Apart from using AI in customer-facing roles, Microsoft has also rolled out generative AI coding tools across new product development and existing updates. Around one-third of Microsoft code is now AI-generated, putting the company on par with its fellow tech giant, Google.
With Microsoft due to release its latest quarterly report soon, we’re yet to see how financial performance has been affected by its use of AI.
However, last quarter the company posted a 13% increase in revenue, with CEO Satya Nadella commenting, “From AI infra and platforms to apps, we are innovating across the stack to deliver for our customers.”
Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!
Microsoft has been in a battle for pole position as the world’s most valuable company in recent months. Though often overtaken by Nvidia, it remains several billion dollars ahead of Apple.
More broadly, Microsoft isn’t the only company making productivity gains. Salesforce, a company that has also been going in hard on generative AI and agentic AI, now reports that around 30% of its internal work is handled by artificial intelligence.
Generative artificial intelligence promises to profoundly reshape labor markets, much like previous automation waves did, but with clear differences. Unlike earlier technologies, such as computerization, which primarily automated routine administrative tasks, or robotics, which impacted manual tasks in manufacturing environments, generative AI targets cognitive tasks.
Exemplified by commercially available large language models, or LLMs, such as GPT-4, generative AI can execute complex, nonstandardized functions that are traditionally reliant on human judgment. This includes tasks such as real-time scheduling, dynamic rerouting of transportation resources, and interpreting customer inquiries in logistics operations. Indeed, major global logistics providers already have successfully leveraged generative AI to automate repetitive cognitive tasks, resulting in notable operational efficiencies and responsiveness. Additionally, companies across the logistics sector have utilized generative AI to automate customs documentation, streamline inventory management, and optimize freight networks.
While generative AI can enhance traditional logistics processes—in ways ranging from demand forecasting and supplier negotiations to network design and contract analysis—its broader implications depend on the worker tasks and economic incentives for adoption. Occupations within supply chain and logistics, particularly those involving routine yet cognitively intensive tasks such as billing, payroll, and data entry, are uniquely positioned for potential disruption.
The degree to which an occupation is potentially impacted by generative AI—a scenario referred to as AI exposure—depends on the susceptibility of tasks within occupations to automation or acceleration through generative AI technologies. Scholars have developed various metrics leveraging different methodologies and data sources, including expert assessments, evaluations of patent-related tasks, and analyses of tasks based on required capabilities, to determine the rate of AI exposure.
Our research finds that occupations within the U.S. logistics sector differ markedly in their vulnerability to generative AI, with cognitive-intensive administrative roles exhibiting particularly high vulnerability. For instance, among the more than 200,000 logistics managers—encompassing operations managers, warehouse managers, transportation managers, and similar titles—more than 90 percent of their tasks are susceptible to AI-driven automation, with nearly 100 percent of these classified as core activities, underscoring a substantial displacement risk. By contrast, bus and truck mechanics—a workforce exceeding 70,000 workers—exhibit literally 0 percent task exposure, highlighting the wide gulf in AI automation risk across the logistics ecosystem.
In this essay, we explore the mechanisms that may shape generative AI’s potential to transform the U.S. logistics workforce. In particular, we discuss how the adoption of generative AI might differentially impact various worker roles and explore the potential consequences for workers, including which occupations could have the greatest difficulty transitioning to new roles if disrupted by AI. We conclude by discussing a range of policy implications, outlining how strategic interventions can ensure that the productivity benefits from generative AI adoption translate into widespread economic gains rather than exacerbating U.S. workforce inequalities.
Labor force characteristics of the U.S. logistics sector
As of May 2023, the Transportation and Warehousing industry (or NAICS 48-49, in U.S. Bureau of Labor Statistics terms) is broadly representative of the logistics sector and employs approximately 6.6 million workers in the United States. Since 2010, this sector has experienced consistent employment growth, driven by rising consumer demand and the increasing complexity of supply chains. Table 1 below details the primary subsectors within the Transportation and Warehousing industry, highlighting their distinctive roles in the broader U.S. logistics landscape. (See Table 1.)
Table 1
Within these subsectors, employment is heavily concentrated in a handful of key occupations. Importantly, the potential effects from AI exposure may be concentrated within these occupations. Table 2provides an overview of the largest occupations by employment in the Transportation and Warehousing industry. (See Table 2.)
Table 2
As Table 1 and Table 2 make clear, the U.S. logistics sector comprises a variety of subsectors and occupations—some heterogenous across industries, such as pilots, and others cross-cutting, including stock and material movers. The impact of generative AI on these occupations will vary substantially based on the tasks that primarily make up each job. We turn to this AI exposure next.
AI exposure in the U.S. logistics workforce
Our analysis begins with a detailed list of tasks performed by workers in logistics occupations, captured through the Occupational Information Network, or O*NET. O*NET comprehensively documents these tasks, including their relative importance to job performance, how frequently each task is performed, and whether each task is “core” or “supplemental.” Core tasks are central to an occupation’s primary responsibilities, while supplemental tasks play a supportive, yet less fundamental, role.
The distinction between automating core versus supplemental tasks is crucial, as research underscores that automation of core tasks typically has more significant implications for occupational stability and employment outcomes. Similarly, as various scholars highlight, task frequency is an important differentiator of AI exposure across occupational tasks. Task frequency directly influences the cumulative economic returns of automation by highlighting repetitive tasks that yield substantial productivity gains.
Explicitly integrating task frequency alongside task importance and the core-supplemental classification can therefore yield a more robust, economically meaningful assessment of AI exposure. Figure 1 below illustrates the relationship between task importance and task frequency within occupations, displaying that task frequency exhibits substantial independent variation—though also showing it positively correlates with importance, too. (See Figure 1.)
Figure 1
We utilize our frequency-weighted AI exposure metric to calculate AI exposure by weighting tasks according to the estimated number of times they are performed annually, based on frequency data reported in O*NET. This approach provides a more economically grounded and practically relevant measure, aligning closely with real-world automation incentives in U.S. supply chain and logistics occupations. This refined exposure metric can better identify the tasks and roles most susceptible to generative AI-driven changes, thus supporting proactive policy formulation and strategic workforce planning.
Logistics sector occupational impacts of generative AI
Understanding generative AI’s occupational impacts within the logistics sector requires a detailed analysis across specific roles and transportation modes. Within the logistics sector, there are specific occupational categories, including truck and water transportation, support activities for air transportation, and warehousing and storage, among others. For each of these categories, we calculated AI exposure scores, derived from our modified AI exposure methodology described above, and weighted them by task frequency. (See Table 3.)
Table 3
What clearly emerges from Table 3 is that different transportation modes within the U.S. logistics sector—for example, freight trucking, water transportation, and pipeline transportation—vary significantly in their exposure to generative AI disruptions. Freight trucking, characterized by dynamic routing, real-time scheduling, and frequent documentation requirements, shows particularly high potential for generative AI exposure, especially in administrative roles. By contrast, water and pipeline transportation, which involve more specialized manual tasks and fewer frequently repeated cognitive activities, exhibit relatively lower immediate susceptibility—though predictive maintenance and monitoring tasks remain promising AI applications.
Notably, rail, freight, and logistics services—and particularly freight transportation arrangement services and customs brokering services—exhibit the highest exposure of any logistics industry analyzed here. With more than 75 percent of tasks potentially decreasing in duration by 50 percent or more, this industry could see a reduction in total employment over the next decade while maintaining its productivity level and increasing its profitability. Although some occupations might maintain employment levels after adopting productivity-enhancing technological innovations, widespread disruption at the sector level inevitably leads to workforce displacement.
Consequences of AI exposure for logistics-sector workers
While exposure to AI offers significant benefits to firms in terms of reduced labor costs and improved productivity, the downside for workers—particularly those in vulnerable roles—includes potential job displacement and wage suppression. Figure 2below introduces this dimension by showing the variation in average annual earnings versus LLM exposure by industry. (See Figure 2.)
Figure 2
The blue bubbles in Figure 2 reveal a striking bifurcation within the trucking and ground-freight subsector. On one hand, truck drivers—by far the largest occupation, employing more than 1 million people—earn relatively modest wages and face relatively low AI exposure, implying limited technical potential or economic incentive for automation. On the other hand, a much smaller but substantially higher-earning group of logistics managers—such as operations, warehouse, and transportation managers, totaling more than 100,000 employees in this sector alone—sits at the very top of the LLM-exposure index, with more than 90 percent of their tasks susceptible to AI. This contrast underscores how, within the same subsector, firms may eschew automating low-wage, low-exposure roles yet aggressively target the high-wage, high-exposure managerial positions for AI-driven productivity gains—along with the attendant displacement of jobs or wage-pressure risks.
Much of our analysis so far has focused on the direct automation of workplace tasks, but indirect exposure to AI also merits consideration. Take, for example, administrative assistants who manage logistics and scheduling for warehouse operations. Even if warehouse loaders have limited direct AI exposure, administrative staff increasingly use AI-driven software to optimize inventory placement and shipping sequences. Predictive models can ensure that frequently ordered items are placed near loading docks, significantly reducing retrieval time for warehouse loaders. Thus, productivity in occupations without direct task automation can improve substantially due to spillovers introduced elsewhere.
This reveals two lenses through which to evaluate AI automation within an industry: one where workers’ tasks are automated to decrease labor hours demanded of that worker type (or to free up worker capacity), and one where the quality of task execution is improved by automation, allowing for efficiency spillovers.
Occupation-specific vulnerabilities to AI adoption
Now that we have showcased the variation across logistics subsectors and occupations of AI exposure, we turn to how specific occupations may be vulnerable to automation and thus to AI adoption. Deeper dives into two occupations that are highly exposed to AI within the logistics sector help illuminate potential vulnerabilities.
Two highly exposed occupations that are heavily represented in logistics industries are customer service representatives and dispatchers (except police, fire, and ambulance). These two occupations each score the maximum value in our exposure index—100 percent—meaning that all their typical tasks are exposed to LLM-powered solutions. This makes these occupations highly vulnerable to job displacement or wage losses as a result of AI adoption.
Let’s turn first to the example of customer service representatives.
Logistics customer service representatives
The typical earnings of customer service representatives—an annual median of $39,680 in 2023—are low, compared to peer occupations requiring similar distributions of skill, knowledge, abilities, and work activities. If displaced, these workers may be relatively more likely to transition to jobs with wages comparable to those they are accustomed to earning. Yet, to the extent that a displacement shock from AI affects similar occupations, these transition options could simultaneously become restricted, placing significant pressure on wages and employment for disrupted customer service representatives. (See Figure 3.)
Figure 3
Job displacement in customer service seems likely, as core tasks such as “confer with customers by telephone or in person to provide information about products or services, take or enter orders, cancel accounts, or obtain details of complaints” and “keep records of customer interactions or transactions, recording details of inquiries, complaints, or comments, as well as actions taken”were already being automated for telephone and virtual support lines prior to the increased availability and accessibility of generative AI. Now, AI automation in customer service is simply a matter of paying for one of the many available services.
Unfortunately for these workers, their next potential occupation may not be a safe harbor from displacement as a result of future automation. When calculating the average exposure level (weighted by historical occupational transition shares, or the expected level of exposure across the set of likely alternative positions), we find that the statistical worker in fields similar to customer service across the U.S. economy has a 95.22 percent AI exposure. Though this is less than the 100 percent score we found for customer service representatives, it is not low enough that these other workers should feel insulated from further disruption. Even if these workers deviate from traditional career pathways, similar occupations tend to be more exposed than the economywide median. (See Figure 4.)
Figure 4
The widespread adoption of automation solutions may therefore lead to disruptions across this occupational cluster. If productivity gains do not lead to compensating demand for labor (fewer workers per task, but a greater volume of tasks to be performed), then we may expect downward pressure on wages, as workers compete for a limited number of job opportunities.
Logistics dispatchers
Our second example occupation—dispatchers (except police, fire, and ambulance)—earned a median wage of $46,860 in 2023. Dispatchers are paid comparable wages to workers in peer occupations requiring similar distributions of skill, knowledge, abilities, and work activities. (See Figure 5.)
Figure 5
We might expect workers transitioning involuntarily (due, for example, to a technological shock) from employment in one occupation to enter a new occupation at or below their current percentile of earnings. In other words, we expect a worker at the 90th percentile in one occupation to be offered a position with pay below the 90th percentile in their new occupation due to their comparatively lower experience in their new role. If so, then Figure 5 suggests that displaced workers at the top of the income distribution in this particular occupation may face significant challenges in maintaining their level of earnings after an occupational transition.
Despite typically earning more than customer service representatives, dispatchers are already subject to automation. This makes their replacement by AI just as technically feasible—and even more incentivized from an employer’s standpoint, as the potential wage savings are greater.
Additionally, the quality of dispatchers’ work has greater influence on the productivity of their colleagues within firms, compared to customer service representatives. This means that quality is an important metric to consider. If generative AI technologies produce worse results than a human, those effects will be magnified, causing a ripple effect throughout organizations. Likewise, if generative AI proves superior to a human dispatcher, then the potential savings may exceed the compensation of the displaced workers, as affiliated workers gain efficiency benefits from improved coordination.
Importantly, the tolerance for AI-committed errors will vary by setting, depending on the cost of the failure in relation to the nature of the error committed, the existing rate of human error, and the feasibility of identifying and perhaps correcting errors. This variance could be across industries—for instance, a low tolerance for error in piloting aircraft versus in warehousing—as well as by the position of the task in the value chain. Errors in delivery, for example, are potentially costly both from lost productivity and because they affect the customer experience, while errors in optimal storage might affect costs through productivity alone.
Like customer service representatives, dispatchers also face challenges in moving to jobs less exposed to AI, though dispatchers are slightly better off than customer service representatives. Our research indicates that similar jobs to dispatchers also are highly exposed to AI, though not as exposed as 100 percent, which is the score we found for dispatchers. (See Figure 6.)
Figure 6
If diverse industries adopt automation at varying rates, displaced workers may be displaced multiple times over the course of a few years. Each time, competition for similar jobs would become steeper, as a growing number of workers fight to the bottom of the earnings distribution. Much worse would be a scenario where many similar industries adopt automation technologies simultaneously.
Economic and operational incentives for logistics firms to adopt generative AI
A firm’s decision to adopt generative AI in logistics, as in any other industry, is ultimately driven by economic and operational incentives. Logistics operations are typically labor-intensive, involving substantial labor costs associated with moving and managing goods. Consequently, there are considerable economic incentives to deploy AI to automate high-frequency, routine tasks—such as documentation, tracking, and inventory management—particularly in sectors where labor constitutes a significant portion of operational expenses.
Labor costs can further shape AI adoption decisions. Higher-wage roles in logistics, such as transportation managers or supply chain analysts, often involve cognitive tasks highly suited to AI tools. Automating or augmenting these tasks can deliver substantial cost savings and productivity improvements for firms. Conversely, lower-wage roles typically offer fewer immediate incentives for AI adoption, not only because these positions often entail tasks less amenable to automation but also due to limited immediate economic returns. Furthermore, retaining employees in these lower-wage roles could enhance overall efficiency through productivity spillovers, reflecting traditional capital-labor complementarities.
A relevant quantitative example demonstrates the varied impacts of AI assistance on customer service agents’ handling times, depending on task complexity. The researchers found that AI significantly improved efficiency for moderately uncommon problems, suggesting substantial benefits through reduced labor costs per interaction. Conversely, AI had minimal impact on very routine or extremely rare problems, implying potential scenarios where AI implementation and maintenance costs might outweigh benefits for firms.
The return on investment for AI adoption in logistics, therefore, similarly depends on specific task characteristics and the corresponding efficiency improvements that AI might realistically achieve in the near term, considering uncertainty about the long-term technical potential of AI.
An important obstacle to AI adoption is resistance from inside or outside of firms. Internally, the resistance might come from decision-makers themselves, who could be put in a position to impact the scope and amount of work available to them by choosing whether to adopt AI. Similarly, lower-wage workers are more likely to be represented by a labor union that will advocate on their behalf against any possible adverse impact of AI adoption.
Policy and regulatory considerations of generative AI adoption
Policymakers face significant challenges in managing U.S. labor market disruptions arising from generative AI. Effective regulatory frameworks must balance the promotion of innovation and productivity gains with safeguarding employment and ensuring equitable outcomes.
One option policymakers might consider is establishing comprehensive worker-transition policies, including robust reskilling and upskilling initiatives. While the existing literature highlights the potential importance of targeted training programs to equip displaced workers with skills aligned with emerging labor market needs, recent analyses suggest caution in this area, noting mixed evidence regarding the effectiveness of traditional retraining efforts. Consequently, policymakers should ensure these retraining initiatives are thoughtfully designed, evidence-based, and specifically adapted to address the unique challenges posed by AI-driven displacement.
Additionally, transparency standards for reporting employment changes due to AI adoption are essential. Developing standardized frameworks can facilitate the systematic collection and dissemination of data on job losses, job creation, and shifts in occupational demand resulting from generative AI. Such transparency helps policymakers and stakeholders alike monitor the impacts of generative AI more accurately, enabling timely and informed interventions.
Relatedly, policymakers can also leverage existing U.S. Bureau of Labor Statistics data collection methods or develop new indicators to identify early warning signs of occupational disruptions from AI. Recent literature highlights the utility of analyzing job postings to detect shifts in skill demands and potential vulnerabilities related to AI exposure.
Additionally, drawing on the experience of existing programs—such as the Trade Adjustment Assistance Program, a federal initiative administered by the U.S. Department of Labor that provides training, reemployment services, and income support to workers displaced by increased importing of goods—can offer valuable insights for designing AI-specific programs. This approach would be particularly useful for creating incentives that encourage businesses to transparently share employment-impact data without inadvertently motivating executives or shareholders to accelerate AI adoption for short-term gains.
Moreover, policies incentivizing firms toward labor-complementing rather than labor-replacing AI technologies could help mitigate employment losses. Worker decision rights around technology adoption could help drive such incentives. Encouraging investment in AI technologies that enhance human productivity, such as predictive analytics and workflow automation tools, rather than entirely substituting human roles, can help maintain employment levels while driving productivity improvements.
Possible policy levers in this area range from tax credits focusing on specific types of technological investments, emphasizing labor-complementing productivity gains, to reskilling programs (for example, accelerating transitions into new occupations for at-risk workers or helping workers gain skills that can help widen their scope of potential transition options to increase their robustness to technological disruption), as well as specific tax credits contingent on payroll targets, such as linking capital investment to wage enhancement.
Regulatory policies must also consider the competitive dynamics introduced by generative AI. Data are increasingly recognized as a critical complementary asset for leveraging generative AI effectively. Policies promoting equitable access to relevant data, such as public data assets or interoperability standards, and those that mitigate monopolistic tendencies in data ownership and processing, such as facilitating market-based purchasing of computing resources, can ensure broader economic benefits and prevent entrenched advantages among incumbent firms. Ensuring a competitive landscape by AI adopters could enable downward pressure on prices, thus exerting an upward force on output and hence labor.
Finally, strategic regional and sector-specific economic planning, informed by detailed analyses of occupation-specific AI exposure, is essential. Policymakers should proactively identify vulnerable communities and occupations, facilitating targeted support through direct interventions, economic diversification initiatives, and stimulus for job creation in sectors less exposed to AI disruption. The geographic concentration of employment in logistics-related occupations makes regional strategies more urgent. For instance, BLS data show a much higher concentration of transportation, storage, and distribution managers in some states versus others, which could make those local labor markets more vulnerable to disruption in occupations with correlated technology exposure.
Integrating these policy considerations can help governments and stakeholders navigate the complex labor market dynamics posed by generative AI, fostering inclusive growth and workforce resilience in the U.S. logistics sector.
Conclusion
This essay provides an assessment of generative AI’s potential impacts within the U.S. logistics and supply chain sector, highlighting key quantitative findings and clarifying implications for policy and industry stakeholders.
Our analysis demonstrates that roles within the U.S. logistics sector exhibit starkly divergent AI-automation risk profiles. Among the more than 200,000 logistics managers—including operations managers, warehouse managers, transportation managers, and similar titles—more than 90 percent of tasks are susceptible to AI-driven automation, with virtually all of those classified as core activities, signaling a substantial displacement risk. In contrast, bus and truck mechanics—a workforce exceeding 70,000—face virtually 0 percent task exposure, underscoring the wide gulf in technical potential and economic incentive for automation across the logistics ecosystem.
These findings emphasize the necessity of nuanced, role-specific workforce interventions and strategic adoption of generative AI technologies. Importantly, individual sectors within the logistics industry also exhibit marked differences in AI exposure. Freight transportation, for example, demonstrates particularly high vulnerability, whereas warehousing and storage faces relatively lower susceptibility. Mobility within occupational clusters and across salary ranges further complicates worker transitions, necessitating targeted policy responses that address potential dislocation and ensure sustainable career pathways.
Policymakers must also consider measures to address long-term earnings losses for displaced workers. Comprehensive policies, including wage insurance, transitional income support, and targeted reskilling and upskilling initiatives, are crucial to mitigate economic hardships. Future research should continue tracking employment outcomes and productivity changes post-AI adoption to refine these policy strategies and improve their effectiveness.
About the authors
Christophe Combemale is an assistant research professor at Carnegie Mellon University’s Department of Engineering and Public Policy, where his research focuses on technological impacts on workforce skills and organizational structures. His recent work explores labor implications of generative AI and industrial transitions across multiple sectors. He is also CEO and principal partner of Valdos Consulting, specializing in techno-economic modeling to support market, technology, and workforce strategy.
Laurence Ales is a professor of economics at Carnegie Mellon University’s Tepper School of Business. He received a Ph.D. in economics from the University of Minnesota in 2008 and a B.S. in physics from the University of Rome Tor Vergata. His research focuses on the design of tax policy and on the labor implications of technological change.
Dustin Ferrone is a senior engineer at Valdos Consulting, providing expertise in complex systems analysis, labor and service supply chains, and scalable analytical software solutions. With a background in systems engineering and extensive industry experience, Ferrone advises both public- and private-sector clients on managing workforce transitions and technological integration.
Andrew Barber is a senior economist at Valdos Consulting, specializing in economic modeling, labor market analysis, and the impact of emerging technologies, such as generative AI. His research evaluates public-sector initiatives, workforce preparedness, and incentive-compatible policy solutions, supporting government and industry efforts to navigate technological disruptions.
Did you find this content informative and engaging? Get updates and stay in tune with U.S. economic inequality and growth!
Stay updated on our latest research