According to an article published on Salesforce's blog, agentic AI helps workers perform tasks they could not do before, but at the same time raises the question of whether it could also weaken existing skills. The researchers examined how knowledge workers use AI to expand their capabilities while preserving the human expertise and judgment that make their work valuable. The findings point to a paradox: a senior designer-engineer interviewed who had relied heavily on AI for months recounted freezing when his partner asked him for help writing and felt he had forgotten how to write; in contrast, an engineering architect described the opposite experience, in which things he wanted to do before and could not became possible for him now. The researchers noted that the core question is not simply whether AI increases productivity, but how to design interactions with it to get better at work and gain the lift without losing the skill.
Scope of the Salesforce Teams' Research
Salesforce's research teams examined human-AI collaboration to understand how to help people thrive in an AI work environment, not just by improving productivity, but by enabling creativity and professional judgment to lead meaningful decisions. As part of the research, 17 employees at Salesforce across 11 different roles were interviewed, including engineers, lawyers, and designers. In addition, micro-interviews were conducted with 2,714 Salesforce community participants (Salesforce Trailblazers) at the Dreamforce conference, 47 academic papers were reviewed, and more than 3,000 real AI work samples were analyzed.
The researchers emphasized that the individuals studied did not merely engage in textual chats with AI tools, but built systems: they embedded their own standards and judgment into reusable AI skills, agents, and reviewers. From the initial research, one central theme emerged: many experts resist passive AI use by engineering deliberate skepticism into their workflows. Although they expressed high trust in AI, that trust was earned through deliberate routines and repeatable processes designed to steer and verify outputs. According to the study, even the heaviest users of coding agents insisted on keeping the "last mile" of human judgment in their own hands.
The Automation Trap and Productive Friction in Workflows
According to the researchers, when skepticism is not built into the process and the system operates without meaningful oversight, AI can encourage passive acceptance rather than active judgment. This phenomenon is known as automation bias, where humans become overly reliant on automated recommendations and outputs. The researchers describe this as the constant repetition of the question "is this okay?", which ultimately turns the answer "yes" from a conscious decision into an automatic reflex. This is the shadow side of the paradox: the frictionless automation that makes an expert faster can bypass the judgment that makes that person trustworthy.
Experts interviewed in the study encountered AI failures that prompted them to build more rigorous checks into their workflows. Among the examples cited in the study: a model describing the contents of a file it had never opened, and an automation system reporting that a task was complete even though it had not run at all. The response of those experts was not to reduce trust in AI, but to embed skepticism directly into the system so it would not rely on willpower alone to counter automation bias.
The researchers define this approach as a core design move: designing for productive friction. Building on the concept of mindful friction, which introduces intentional pauses to help prevent harm, productive friction goes a step further and aims to design AI interactions that actively strengthen human judgment, creativity, and expertise, rather than merely removing obstacles or completely automating decisions. For example, a technical writer interviewed in the study caught an AI model pretending to call a live repository of her professional context layer (product information, vocabulary, etc.). Following this, she built a mandatory verification loop asking "are you sure?", which requires the model to pause and confirm its level of certainty before proceeding.
Eight Practices for Preserving Human Judgment and Expertise
The researchers identified eight emerging practices designed to keep human expertise and judgment at the center of AI systems:
-
Enable the ability to apply human judgment where it matters: Building opportunities for users to exercise their expertise at key decision points. This includes domain-specific agents or skills that recognize when a decision requires human judgment rather than automation. For example, an AI system can pause at a critical decision point and ask the user whether the recommendation feels right, and if something feels wrong, help them explore why before proceeding.
-
Human goes first: Preserving the user's independent thinking before the AI provides a solution. This approach may include a staged reveal, where the AI's response is withheld until the user completes their initial assessment or ideation. Another approach is using AI as an editor rather than a generator, which preserves the user's original thinking and creativity. Additionally, users can define intent, specifications, and a work plan before asking the AI to build the deliverable.
-
Engineer scrutiny: Embedding verification directly within AI workflows rather than relying solely on human vigilance to catch mistakes. One participant in the study built an audit log for her coding agent sessions. Every time she corrects an error, questions an AI claim or decision, or modifies a deliverable the agent built, the system records it. She regularly reviews the logs to identify patterns and adjust the agent's autonomy, instructions, and error-checking processes.
-
Get a second opinion: Embedding independent perspectives that challenge assumptions and surface gaps. Rather than relying on a single model or tool, practitioners can use multiple agents to evaluate the same task, compare conclusions, or debate different approaches. A designer-engineer interviewed built an "AI council" consisting of expert personas from different fields, such as security, engineering, and inclusive design. The agents debate potential solutions and report back on areas of disagreement, which helps counter overly agreeable AI responses and sharpens understanding of the problem.
-
Use AI to engage complexity, not collapse it: Resisting the oversimplification of issues and helping explore nuance and trade-offs. Many interviewees shared that AI can make complex issues seem simpler than they are, delivering confident answers that overlook important context. Rather than doing all the sensemaking for us, AI can present multiple plausible interpretations of an ambiguous question side by side and explain how different priorities might alter the recommendation.
-
Own context creation and quality standards: Helping users define what counts as "good" rather than inheriting those standards from AI outputs. This can be done through shared team knowledge bases ("second brains"), where workers collaboratively build and maintain the context relied upon by AI agents, as well as by creating high-quality test sets and evaluation metrics to assess AI outputs against agreed-upon standards.
-
Prevent dependency: Protecting user autonomy and avoiding design that creates unnecessary attachment. Interviewees described sessions where the ability to produce outputs quickly made spending more time and energy than planned all too easy. To counter this, AI experiences can be designed to provide natural stopping points that signal the completion of a task and invite the user to move on, rather than ending every response with another follow-up question.
-
Match oversight to the stakes: Adjusting scrutiny and human involvement to the consequences of a potential error. A rough internal draft may require only a quick check, whereas customer-facing, executive-level, regulated, or irreversible decisions may demand more rigorous verification or explicit human approval.
Recommendations for Team Leaders and Product Builders
At the conclusion of the study, the researchers offer practical steps for action: For leaders of teams adopting AI, it is recommended to identify the quality assurance rituals that top-performing employees have already built, often without giving them a formal name. The suggested question to ask them is: "If someone else tried to scale a slice of your work, what check would they miss because they don't have your expertise?" This answer provides a foundation for scaling best practices through design and implementation decisions.
For AI product builders, the researchers recommend establishing human thriving as an explicit goal alongside quality and productivity. According to the article, AI can serve as a genuine force multiplier for professional skill, but only when workflows are designed to protect human agency, introduce a little productive friction, and treat craftsmanship as a core part of the work rather than just a nice-to-have. The researchers conclude that those who thrive are not necessarily those who trust AI the most, but those who consciously decide where to keep themselves at the helm. The researchers noted that they will continue to expand on this topic in future publications.