在时钟上：工作场所的代理人工智能

On the clock: Agentic AI in the workplace 18September 2025 Key takeaways •Agentic AI refers to autonomous, action-taking AI systems that have the potential to automate complex business processes.This evolution shifts generative AI from a reactive tool to a proactive and goal-driven virtual collaborator. •While enterprise adoption of agentic AI remains in very early stages, momentum is building. By the end of 2025, InternationalData Corporation predicts half of all organizations will use enterprise agents tailored to specific business functions.•As AI technology improves at a breakneck speed, it's challenging to predict the pace and magnitude of the agentic AI adoptioncycle. Yet the long-term potential is undeniable. In fact, BofA Global Research predicts that agentic AI's total addressable market(TAM) will reach $155 billion by 2030. Agentic AI 101 What is it?Agentic AI refers to AI systems that can achieve a set goal or task with limited supervision.1Unlike generative AI, which creates content in response to a user’s prompt or request, agentic AI is made up of agents that can autonomously make decisions orperform tasks, utilizing reasoning capability to choose the correct tools for the job. For example, an AI agent could fill out forms,make dining reservations, or book your travel arrangements–by itself. Does this topic seem familiar? For more, look back at ourJanuary publication,The new wave: Agentic AI. Agentic AI architectureAgentic AI is made up of a number of agent layers. There are common components that help structure each agent layer, and these components remain very similar across different types of agentic AI. According to BofA Global Research, as agentsbecome more advanced, the number of existing agent layers is likely to expand, however, the core components are anticipated toremain unchanged (Exhibit 1). According to International Data Corporation (IDC), agents operate by combining three distinct layers of capability. First, we haveinteraction–where agents interact with their environment to receive information about a required goal. They monitor theenvironment to receive feedback about progress toward the goal and to provide feedback to the requester. Second isplanning.Given a goal and knowledge about its current environment, an agent creates a plan (a set of actions) that will enable it to achieveits goal. Lastly, there’saction, which may involve sending messages, invoking external systems, updating data, and setting goalsfor other agents (to enable "multi-agent collaboration"). For example, consider what it would look like if a large enterprise leveraged agentic AI to manage IT incidents across itsinfrastructure–servers, networks, and applications.Interactionindicates that the agent would consistently monitor system logs,performance metrics, user reports, etc.–to look for anomalies. Once identifying an issue, it would then receive a goal: restoreperformance to normal levels. It could also communicate with IT staff, providing updates and receiving feedback. Forplanning,based on the agent’s understanding of its environment, it would formulate a plan for issue resolution, which it would thenexecute in theactionstage. If the issue required broader coordination, it could also set goals for other agents–to successfullyensure multi-agent collaboration. Summary of the various agent layers and their common components Agentic AI in the workplace Adoption requires elevated workflow orchestrationAs illustrated inExhibit 2, increased enterprise adoption of agentic AI would elevate the need for workflow orchestration and task management, seeing as roles and tasks in the workplace would be distributed between humans, robots, and AI systems,depending on skills required (for more on this topic, readAI dictionary, part 2: The next generation). Ultimately, chatbots andcopilots (i.e., generative AI that acts as a“virtual assistant”by providing support, suggestions and automating tasks) willcomplement agents as they develop and mature. Exhibit2:Agentic AI will increase the need for workflow orchestration and task management, choosing between humans,robots, and AI systems to complete tasks independently or collaboratively based on competenciesIllustration of Human / Machine skills and teaming: human, agentic AI-human, agentic AI, and robots How much more productive could AI agents make humans?The potential use cases for agentic AI are vast, spanning from an increase in productivity for marketing professionals to software developers. In fact, a recent study conducted by the Massachusetts Institute of Technology (MIT) evaluated potentialproductivity increases for marketing professionals via the usage of AI agents.2Interestingly, the study found that individualsplaced on teams collaborating with AI agents were 60% more productive and created higher-quality ad copy than the individualsplaced on human-human teams (with the exception of images, suggesting AI agents require fine-tuning for multimodalwor