Create an Onboarding Plan for AI Agents

Most executives believe that the big challenge in adopting agentic AI is figuring out how to adapt to a new and important technology. But in fact it’s primarily about managing work. Executives have been slow to recognize this in part because of AI’s still nascent capabilities. According to a recently published and widely cited paper by Anthropic’s head economist, Peter McCrory, and his colleague, Maxim Massenkoff, a yawning gap exists between what AI can theoretically do and how it is actually deployed, even in areas undergoing profound transformation. Most estimates suggest that 94% of the tasks in computer- and math-related occupations, for example, are displaceable by generative AI, but the authors write that Anthropic’s current offerings only cover a third of those tasks. But a far greater gap exists on the human side of the equation. Multiple studies released around this year’s World Economic Forum (by Deloitte and McKinsey, for example) indicated that less than 10% of companies felt they were making substantial progress in designing effective human-machine interactions.To capture AI’s near-term benefits and prepare for its expanding impact, organizations need to integrate it into existing HR processes and make its role clear to employees. Below are six ideas I’ve developed working with companies to realize that objective. They come from a range of industries, including banking, consumer products, logistics, and life sciences, and are widely recognized as leaders in their respective industries. Give every AI agent a job description. As agentic AI penetrates organizations, teams will consist of human and agentic AI “colleagues,” and all of them will need job descriptions—a tried-and-true way of specifying a worker’s responsibilities, decision rights, and role in relevant processes. Crafting a job description for every AI agent will oblige their managers to be deliberate about allocating responsibilities across human and agentic colleagues. As you craft those agents’ job descriptions, ask yourself these questions: What is it responsible for—and not responsible for? (Vague mandates such as “optimize” or “improve efficiency” are a recipe for trouble. Just as people do, agents perform best when objectives are explicit.) What are the boundaries of its decision rights? What are its authorities? When must it seek input or approval from colleagues or superiors? Design the agent to address pain points for human colleagues. Many observers have pointed out that agentic AI promises to improve the quality of knowledge workers’ work lives. Just as automation helped eliminate work in industries that were, as the saying goes, “dirty, dark and dangerous,” such as manufacturing, metals processing, and mining, AI can help eliminate work that is “dull, dispiriting, and deterministic.” Whether by reducing tedious steps or automating the dreariest elements of a job, designing jobs for AI agents in this manner gives employees a reason to adopt AI and encourages them to work around its limitations by grounding it in their day-to-day experience. Evaluate every AI agent on a regular cycle. AI agents need measurable performance metrics for the actual process outcomes that arise from their actions. The metrics have to be explicit and allow the broad evaluation of performance, going beyond qualities such as accuracy and ease of use to include timeliness and reliability. The existence of such metrics will help reassure human teammates that their algorithmic colleagues are being held to standards, too. Moreover, such metrics will help improve training regimes by revealing areas for improvement. Much as performance reviews inform professional development plans for employees, agentic teammates should benefit from a cycle in which feedback informs learning. Without metrics, managers can’t distinguish between acceptable variation and real failure. Give every AI agent a human supervisor. Observers have begun to tout the capacity of AI agents to “orchestrate” the activities of many discrete agents. That may speed actions and make their work more precise. But who will oversee the orchestrators? The need for human oversight undoubtedly remains. Every generation of AI has demonstrated some propensity to hallucinate. Instances are likely to wane as AI improves, but the stakes will mount as AI expands into professions deemed to be significantly exposed. Moreover, organizations will remain accountable for any results generated by AI it employs, so a sentient decision-maker must be held accountable for how it’s trained, how it integrates with processes, and how it interacts with other human and agentic teammates. Regulators, legislators and the courts are certain to insist on that. Hire AI agents as interns and make them earn full-time status. Precocious interns can help you and your colleagues. You can ask a lot of them, but you also need to provide them with clear training, guidance, and structure. New AI agents are analogous. They’ve been trained (that is, taught a lot) about the basic concepts of the task they’re about to undertake. But they lack meaningful practical experience. They don’t have contextual intelligence about your company’s culture, values, and strategy, and about the workings of specific processes or markets. You hire interns to let them gain some of that experience and to see how they perform. Treat AI agents the same way. Don’t hire them on a full-time basis until they demonstrate they have the ability to perform to their job descriptions within the performance parameters established. No agentic AI should be “hired” based on the extent of its training or promises of what it will eventually accomplish. Only a proven agentic AI should be made a permanent part of an ongoing process and introduced to other parts of an organization. Give each AI agent a name. Don’t do this to humanize the agent, but to make its role discussable. When people say, “This decision came from AI,” it is easy for colleagues to ignore agents’ roles as the equivalent of teammates. Their sense of personal responsibility for the process’s outcome can evaporate. And there may be multiple agents involved. By breaking down responsibilities and assigning them to AI agents as you would to human employees, and by giving each agent a clear identity (as in the well-known cases of Watson, Alexa, and Siri), their roles become easier for people to understand. . . . Embedding agentic AI in large organizations will take more than demonstrating a business case. It will require rethinking the way work is managed and developing a pathway to achieve the transitions required. The quickest means to achieving that end is to use mechanisms for managing work that are familiar to executives, managers, and employees alike. Extending those tools to incorporate the burgeoning numbers of virtual agentic workers will ease a transition made daunting by the breadth and depth of change involved.

Create an Onboarding Plan for AI Agents

Summary.

Partner Center