Research: Why Some Junior Employees Work Well with AI-and Others Don’t

Entry-level employees are on the front lines of a rapid shift in knowledge work. These jobs used to be full of tasks that created an onramp into an industry and helped junior employees start building expertise. But now, many of these tasks are at risk of being delegated to AI-powered workflows, which are increasingly capable of handling analytical, information-intensive assignments. And as AI continues to improve, research has shown it is quickly resetting the baseline, with the standard for acceptable output rising as models improve. In this context, organizations need to understand how individuals create value beyond the AI baseline in real organizational settings, particularly in higher-stakes professional work requiring judgment, domain expertise, and decision-ready outputs. That raises two big questions: 1) What enables early career employees to add value in AI‑enabled workflows? And 2) how can those capabilities be intentionally developed in ways that sustain continuous improvement as AI continues to advance? Leaders need clear, explicit answers to both. If they don’t, they risk undermining the return on both their organization’s AI investments and talent pipelines. To help answer these questions, KPMG and the University of Texas at Austin conducted a large-scale field study with 523 U.S.-based early-career professionals at KPMG with tenures of 18 months or less to better understand how to create value in this context. The research was designed as a field study, where participants completed business-specific tasks using an AI agent built for the domain. These tasks were designed by KPMG senior experts to mirror the work early-career professionals perform in practice. They are simplified versions of client engagement tasks that require participants to analyze business information, exercise professional judgment, and develop recommendations. To create a baseline, we began the field study with AI agents performing the work alone. Then, we asked participants to do the same work using the AI agent. Participants’ deliverables were evaluated using detailed grading rubrics developed by KPMG domain experts to reflect real-world performance expectations, and their performance was benchmarked against the AI-only baseline. We also assessed participants’ foundational skills including domain knowledge, critical thinking, and AI literacy, and analyzed their interactions with the AI to understand how they applied these skills while completing the tasks. The goal here was to identify who adds value beyond AI, who merely replicates it, and who may inadvertently reduce performance. The results revealed three distinct performance profiles: AI apprentices (24.1%) performed below the AI baseline. AI delegators (25.8%) produced results comparable to the AI baseline. AI amplifiers (50.1%) outperformed the AI baseline. A surprise in the findings was that foundational skills such as critical thinking abilities, AI literacy, and domain knowledge weren’t clear signals for who fell into which group. Rather, how people worked seemed to make the difference. This reveals something important: AI doesn’t standardize work but rather magnifies differences in how people apply their skills. These findings offer an important lesson to companies: They need to rethink how they design entry-level work and measure employee performance. How Amplifiers, Delegators, and Apprentices Work It makes sense why organizations might prioritize for critical thinking, domain knowledge, and AI literacy when considering candidates for entry level jobs. Yet according to our findings, these three qualities alone do not reliably predict who will perform well in entry level jobs in the AI era. Surprisingly, employees with similar levels of foundational skills and knowledge can generate dramatically different outcomes when working with the same AI agent. AI Apprentices: Capable, but failing to translate skills into value. AI apprentices scored higher than delegators on critical thinking, domain knowledge, and AI literacy and are largely indistinguishable from AI amplifiers on these measures. While they critiqued AI responses, that critique rarely improved the output. Our analysis reveals what is missing in the interactions of apprentices with AI: They did not apply their foundational skills within AI-enabled workflows to guide, evaluate, and refine AI’s output. They often focused on irrelevant issues or guided the AI in unproductive directions. For instance, given the task of reviewing a complex set of business documents and preparing recommendations using AI, the apprentices engaged with AI. But rather than treating it as a way to test their own thinking, they spent most of the conversation requesting, rewriting, or reorganizing information, or pursuing less relevant questions. AI Delegators: Productive on the surface, limited in practice. AI delegators scored lowest on foundational skills, but they were not the weakest performers. Because AI systems already generate competent outputs, delegators proved reasonably productive when performance was judged solely on the quality of the final deliverable. They typically accepted AI outputs with minimal interrogation, added limited reasoning, and contributed little or no additional value. When given the task of reviewing complex documents and preparing recommendations, delegators provided the relevant materials and desired deliverable, then largely accepted the AI’s output with minimal intervention. AI Amplifiers: Able to Turn Capability into Performance. AI amplifiers distinguished themselves not through stronger underlying skills, but through how they combined and applied those skills while working with AI. They actively orchestrated the workflow and framed problems in ways that guided AI toward relevant analyses, anchored work in appropriate domain frameworks, and defined clear evaluation criteria. They interrogated outputs, challenged assumptions, and iteratively refined results. At the same time, they used their understanding of AI to provide precise, task-specific guidance that enabled the system to produce deliverables that met real-world performance standards. In the documents task, amplifiers treated AI as a thought partner, asking it to test assumptions, explain its reasoning, consider alternative explanations, and identify missing evidence before refining the final recommendation. Rethinking Early-Career Potential Our research findings suggest that there is a meaningful opportunity for organizations to rethink entry-level development. Employees who seem to have similar levels of knowledge and skill can produce very different outcomes once AI enters the workflow. But with the right intervention, individuals who struggle today can become strong long-term contributors, particularly if their existing capabilities are better aligned with how work is done alongside AI. Let’s consider each in turn. The AI apprentices represent an interesting paradox: They have the right foundational skills, but they fail to translate them into value. Leaders can unlock this “untapped potential” and turn these employees into high-value contributors by redesigning workflows and training. When employees are given clearer decision rights, practical use cases, and opportunities to apply AI in the flow of work, they can move from simply using AI tools to integrating their capabilities with AI-mediated workflows. AI delegators highlight a different opportunity for growth. Because AI systems already generate competent outputs, these employees can appear productive when evaluation focuses only on final deliverables. Our findings suggest that this pattern may reflect their comparatively weaker foundational capabilities, which may lead them to rely more heavily on AI rather than actively shaping and improving its outputs. The risk is that this behavior remains largely invisible, allowing underperformance to persist while appearing acceptable. Organizations should therefore invest in training that builds both their foundational capabilities and the practical know-how to apply them effectively in human–AI collaboration. AI amplifiers offer organizations a blueprint for what effective human–AI work looks like in practice. Their advantage appears to stem from possessing not only strong domain knowledge, critical thinking, and AI literacy, but also the ability to integrate these capabilities into every stage of working with AI. The opportunity lies in making their behaviors durable and scalable while continuing to stretch their capabilities through more complex, higher stakes work. There’s an opportunity for organizations here. As AI increasingly establishes a high-quality baseline, we expect that human roles will shift from producing answers to directing, evaluating, and extending AI generated work. Training and performance systems that emphasize how employees frame problems, interrogate AI outputs, refine results, and integrate insights into decisions will help more people consistently translate their capabilities into impact. Rethinking the Early Career Experience Enabling early-career professionals is not only about expanding what employees know but also helping them learn how to consistently translate what they know into effective action in AI-enabled workflows. Based on the results of our study, organizations should prioritize contextual, task-based training that focuses on how work actually gets done with AI agents: how employees structure problems, iteratively guide AI systems, critically evaluate outputs, and integrate results into decision-making processes. That means designing AI learning as an ongoing development model rather than a one-time training intervention. KPMG is a test case for this approach. Informed by this ongoing research with UT, the company is trying to apply and scale this approach to build and reinforce “AI amplifier” behaviors across the firm. Rather than a standalone training program, KPMG’s internal AI training program aims for continuous learning that is targeted to increase each individual professional’s AI fluency. Employees begin with skill check to establish their current AI fluency, then move through personalized learning pathways that combine firmwide and function-specific courseware, market activations, podcasts, on-the-job activities, and engagement in a growing AI champions network. Each professional completes a minimum number of hours of structured learning, complemented by experiential activities tailored to their role, level, and proficiency. Early results suggest through this personalization to an individual’s skill gaps, combined with formal learning, real-world application, and peer reinforcement leads to strong capability building. The same principle applies at the functional level, where AI learning must be embedded in the tasks and decisions that shape day-to-day work. For example, at KPMG, simulation-based exercises mirror common client scenarios ranging from reviewing complex tax deliverables to advising on the broader business implications of major market signals and disruptions. In this environment, AI can generate initial drafts and analyses, but human judgment determines how to assess relevance, sequence the work, and decide next steps. This shift is also informing KPMG’s National Intern Training at KPMG Lakehouse, its annual function-specific learning event, which has recently evolved to emphasize these same hands-on, scenario-based experiences, beginning with Audit interns this summer and with plans to extend elements across Tax and Advisory over time. Early indicators suggest that employees who receive personalized, role-based AI development opportunities demonstrate greater confidence, engagement, and effectiveness in applying AI to real-world work. More broadly, our findings can be generalized into three lessons for enterprises preparing early-career employees to maximize value from human-AI collaboration. Rethink human value creation. As AI agents increasingly generate high-quality outputs, we believe that employee contribution will shift from producing answers to directing, evaluating, and extending those outputs. Our research establishes that possessing foundational capabilities including domain knowledge, critical thinking, and AI literacy are necessary but not sufficient for people to create value beyond AI. Therefore, organizational training has to involve employees working not only on paper but also with AI workflows and learning how to steer the AI to create value beyond the AI-only baseline. Make judgment visible early. Organizations should put greater emphasis on surfacing judgment in early career roles. In traditional training environments, strong reasoning often remains implicit. In AI mediated work, KPMG, for example, has begun encouraging practices such as documenting why AI outputs were accepted, modified, or rejected, and articulating the criteria used to evaluate results. This makes human contribution explicit and coachable, rather than assumed. Shift assessment from answers to process. Evaluating employees based solely on the quality of final deliverables is no longer sufficient when AI can independently generate high-quality results. As AI systems become more capable, organizations should instead assess how employees interact with AI throughout the workflow. KPMG is piloting more workflow-based assessment in early-career programs. For example, a functional-specific pilot with audit interns focuses on problem-solving and communication, assessing candidates not only on technical accuracy but on their ability to reason through issues, challenge assumptions, and explain decisions. This approach helps distinguish between those who merely reproduce AI outputs and those who extend them. . . . Taken together, these shifts demand a fundamental transition: from treating AI as a plug-and-play tool for individual tasks to using it as a catalyst to completely rewire how work is executed. The organizations that adapt most effectively won’t just focus on developing ‘AI-literate’ employees; they will build entirely new operating models. As AI agents handle the baseline of early-career knowledge work, differences in value creation will increasingly reflect the extent to which employees are trained to apply their foundational skills within well-designed AI-mediated workflows.