Agentic AI is fundamentally altering the landscape of the modern workplace, granting knowledge workers the ability to accomplish tasks that were previously impossible or prohibitively time-consuming. From automated coding to complex data synthesis, these tools are acting as force multipliers for individual output. However, as the adoption of these systems accelerates, a growing concern has emerged among researchers and industry leaders: while AI may be expanding our capabilities, it could simultaneously be eroding the very human skills and judgment that define professional excellence.
Salesforce researchers are currently deep in the process of exploring this fundamental tension. Their objective is to understand how knowledge workers can leverage AI to amplify their reach without sacrificing the nuance, creativity, and critical judgment that make their work truly valuable. The resulting data points to a striking paradox in the way humans interact with these powerful new systems.
The research highlights two divergent experiences. In one case, a senior designer-engineer, having relied heavily on AI-driven workflows for several months, reported a sense of paralysis when tasked with a writing project from scratch, feeling as though the mental muscles required for the work had atrophied. Conversely, an engineering architect described a sense of liberation, noting that AI allowed him to execute complex projects that were previously beyond his grasp. This dichotomy suggests that the question is not simply whether AI improves productivity, but rather how we consciously design our interactions with these tools to ensure they enhance—rather than replace—human expertise. The central challenge for the modern workforce is clear: how do we harness the "lift" of AI while ensuring we don’t lose the foundational skills that underpin our craft?
Investigating the Future of Human-AI Collaboration
To better understand this dynamic, the Responsible AI and Product Research teams at Salesforce have been conducting an extensive study into human-AI collaboration. The goal is to move beyond simple productivity metrics and instead focus on how these tools can enable creativity and support meaningful decision-making.
The research team employed a multi-faceted approach to gather data. They conducted in-depth interviews with 17 professionals across 11 distinct roles, including engineers, lawyers, and designers. To supplement these qualitative insights, they held micro-interviews with 2,714 Salesforce Trailblazers at the annual Dreamforce event. Furthermore, the team reviewed 47 academic papers and performed a rigorous analysis of more than 3,000 real-world AI work samples.
Crucially, the focus was not on casual users who treat AI as a simple chatbot. The study centered on individuals who are actively building with the technology—people who encode their own personal standards, institutional knowledge, and professional judgment into reusable AI skills, agents, and verification workflows.
A key theme emerging from this initial phase of research is the role of "productive skepticism." Many experts actively resist the temptation to use AI in a passive manner. Instead, they engineer a level of skepticism directly into their workflows. While these individuals maintain high levels of trust in AI, that trust is not given blindly; it is earned through deliberate routines, repeatable processes, and rigorous verification cycles. Even the most prolific users of coding agents within the study insisted on maintaining a "last mile" of human judgment, ensuring that they remain the final arbiter of any output before it is finalized or deployed.
The Trap: When "Yes" Becomes a Reflex
The researchers warn of a significant risk when this skepticism is omitted from the workflow. Without meaningful oversight, AI systems can easily nudge users toward passive acceptance rather than active evaluation—a psychological phenomenon known as "automation bias." When a worker is repeatedly prompted by an AI to confirm a task or validate a suggestion, the decision to say "yes" can eventually cease to be an active, thoughtful choice and instead become an unconscious, reflexive action.
This is the "shadow side" of the productivity paradox. The very same frictionless automation that enables an expert to work faster can simultaneously bypass the critical judgment that makes that expert reliable in the first place. The study found that several experts encountered AI failures—such as a model hallucinating about a file it had never opened or an automation claiming a task was complete when it hadn’t actually run—that served as wake-up calls. These individuals responded by building more robust, deliberate checks into their systems.
Their strategy was not to abandon AI, but to institutionalize skepticism so that the system’s integrity did not depend solely on the user’s willpower or vigilance. Salesforce researchers refer to this as the "design move" of productive friction. While the concept of "mindful friction" involves introducing intentional, small pauses to prevent harm or error, productive friction goes further. It is the practice of designing AI interactions that actively demand the application of human judgment, creativity, and expertise. By embedding these practices into the technology itself, organizations can keep people meaningfully engaged without requiring them to remain in a state of constant, exhausting vigilance.
One technical writer featured in the study provided a practical example of this. After catching an AI pretending to access a live repository of her expert context—including specific product information and company vocabulary—she implemented a mandatory verification loop. The system is now required to pause and confirm its own level of certainty before proceeding. This is not a radical new concept; it is an extension of what seasoned professionals have always done when they know their craft well enough to anticipate where errors typically arise. By placing a check at those specific points, the user effectively creates a "safety net" that allows for faster work without the risk of blind reliance.
Centering Human Expertise and Judgment
Engineering skepticism into workflows is only one of several strategies observed by the researchers. Other emerging practices involve keeping humans firmly in the loop by maintaining control over high-stakes decisions and utilizing AI as a tool to navigate complexity or explore multiple, diverse perspectives. These practices suggest a path forward where AI systems are designed to strengthen human capabilities rather than merely automate tasks to save time.
For leaders currently managing teams that are adopting these technologies, the researchers suggest looking inward. The most effective quality-assurance rituals are often already being performed by top-tier employees, frequently without the workers having formally named them. Managers should engage these experts with specific questions, such as: "If someone else tried to scale a slice of your work, what specific check would they miss because they lack your expertise?" The answers to these questions are invaluable, serving as the blueprint for scaling best practices through better design and implementation.
For those involved in building AI products, the message is equally clear: human thriving must be an explicit objective, treated with the same importance as quality and productivity. AI has the potential to act as a powerful force multiplier for craftsmanship, but this only occurs when the workflow is designed to protect human agency. By incorporating productive friction and treating human expertise as a core component of the process, developers can ensure that the technology serves the user, rather than the other way around.
Ultimately, the professionals who are thriving in this new environment are not necessarily those who trust AI the most. Rather, they are the ones who have made a conscious, purposeful decision about where and how to remain at the helm. As the researchers continue to consolidate these findings, they view these initial practices as essential thought-starters for an industry currently at a crossroads. The goal is to ensure that as we continue to push the boundaries of what is possible with AI, we do so by empowering the people who provide the vision and the judgment that keep our work meaningful.
