The Humane Agentic Enterprise: How To Thrive with AI at Work

Agentic AI is helping us do things we couldn’t do before. But could it also be weakening the skills we already have? To explore this tension, Salesforce researchers are studying how knowledge workers are using AI to expand their capabilities while preserving the human expertise and judgment that make their work valuable. The findings reveal an interesting paradox. A senior designer-engineer we interviewed, who had relied heavily on AI for months, described freezing when his partner asked him to help write something. He felt he’d forgotten how. Meanwhile, an engineering architect described the opposite experience: “Some of the things I wanted to do and couldn’t before,” he said, “I can do now.” The question isn’t simply whether AI makes us more productive. It’s how we design our interactions with AI to make us better at what we do. That tension is the biggest question for anyone working with AI right now: how do you get the lift without losing the skill? What we’re studying Salesforce’s Responsible AI and Product Research teams are studying human-AI collaboration to understand how to help people thrive with AI – not just by making them more productive, but by enabling their creativity to shine and their expert judgment to drive meaningful decisions. To explore we interviewed 17 people at Salesforce across 11 jobs, including engineers, lawyers, and designers. We conducted micro-interviews with 2,714 Salesforce Trailblazers at Dreamforce . We also reviewed 47 academic papers and analyzed more than 3,000 real AI work samples. The people we focused on weren’t just chatting with AI. They were building: encoding their own standards and judgment into reusable AI skills, agents, and reviewers. We’ll be sharing the full research soon, but one key theme emerged from our initial study: Many experts resist passive AI use by engineering skepticism into their workflows. While they expressed high trust in AI, that trust was earned through deliberate routines and repeatable processes designed to steer and verify its outputs. Even the heaviest users of coding agents insisted on maintaining a “last mile” of human judgment as their own. The trap: when “yes” stops being a decision But what happens when skepticism isn’t built into the process? Without meaningful oversight, AI can encourage passive acceptance rather than active judgment. This is known as automation bias, when people become overly reliant on automated outputs and recommendations. Ask someone “is this okay?” often enough, and “yes” stops being a decision and becomes a reflex. That’s the shadow side of the paradox: the same frictionless automation that makes an expert faster can bypass the judgment that makes that person trustworthy. Several experts we interviewed described encountering AI failures that prompted them to build more deliberate checks into their workflows — from a model describing a file it had never opened to an automation reporting that a task was complete when it hadn’t run. Their response wasn’t to trust AI less. It was to build skepticism into the system, so it didn’t depend on anti-automation bias willpower alone. We think of this as the core design move: design for productive friction. Building on our concept of mindful friction , which introduces intentional pauses to help prevent harm, productive friction goes a step further. It’s about designing AI interactions that actively strengthen human judgment, creativity, and expertise, rather than simply removing obstacles or automating decisions. By building these practices into AI systems, we can help people stay meaningfully engaged without relying on constant vigilance. For example, one technical writer, who caught an AI pretending to call a live repository of her expert context layer (product information, vocabulary, etc.), built a mandatory “are you sure?” verification loop into her workflow, ensuring the AI pauses to confirm its own certainty. This isn’t a new concept. It’s what people do when they know their craft well enough to know exactly where AI tends to go wrong, and put a check right in that spot. Engineering skepticism into AI workflows is just one of several practices we observed in our research. Others focus on keeping people actively engaged in their work — from maintaining control over important decisions to using AI to explore different perspectives and navigate complexity. Together, these emerging patterns offer a starting point for thinking about how AI systems can strengthen human capabilities, not just automate tasks. Eight thought starters for centering human expertise and judgment Here are eight emerging practices for keeping human expertise and judgment at the center of AI systems. Enable the ability to apply human judgment where it matters: Build in opportunities for people to apply their expertise at important decision points. This might include domain-specific skills or agents that recognize when a decision requires human judgment rather than automation. For example, an AI system could pause at a critical decision point and ask the user whether a recommendation feels right. If something feels “off,” it could help them explore why before proceeding. Human goes first: Preserve the user’s independent thinking before AI provides a solution. This might include a staged reveal, where the AI’s response is withheld until someone has completed their initial assessment or ideation. Another approach is to use AI as an editor rather than a generator, preserving the user’s original thinking and creativity. This was one of the most common approaches people described at our Dreamforce research wall . Practitioners can also define their intent, specifications, and plan before asking AI to build, ensuring that human direction shapes the work from the start. Engineer scrutiny: Build verification into AI workflows rather than relying only on people to catch mistakes. One practitioner we interviewed built an audit log for her coding agent sessions. Every time she corrects a mistake, questions an AI-generated claim or decision, or modifies something the agent has built, the system records it. She then reviews those logs regularly to identify patterns and adjust the agent’s autonomy, instructions, and error-checking processes. This makes scrutiny a repeatable part of the workflow rather than something that depends on constant human vigilance. Get a second opinion: Embed 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 their conclusions, or debate different approaches. One designer-engineer we interviewed built an “AI council” of expert personas representing different domains, including security, engineering, and inclusive design. The agents debate potential solutions and report back on areas of disagreement, helping the user consider perspectives they might otherwise miss. This approach can help counter overly agreeable AI responses while expanding the user’s understanding of a problem. Use AI to engage complexity, not collapse it: Resist oversimplification and help people explore nuance and trade-offs. Many practitioners we interviewed shared that AI can make complex issues seem simpler than they really are, sometimes producing confident answers that overlook important context. Rather than doing all the sensemaking for us, AI can help us better understand the information, assumptions, and trade-offs behind a decision. For example, an AI system might present multiple reasonable interpretations of an ambiguous question side by side, explaining how different perspectives or priorities could change the recommendation. The goal is to help people navigate complexity and make more informed decisions, rather than simply accepting a single answer. Own context creation and quality standards: Help users define what “good” looks like rather than inheriting those standards from AI-generated outputs. This m