AI agents in action

The podcast "You Can with AI" by KPMG explores the transformative role of AI agents, emphasizing their potential to create new opportunities rather than just improve efficiency. Swami Chandra Sekharan, Global Head of AI and Data Labs at KPMG, outlines a four-part journey in AI development: managing structured and unstructured data, leveraging generative AI and large language models, developing knowledge systems, and implementing agent technology that autonomously executes user-defined goals.

AI agents are defined as entities that understand and fulfill goals through planning, coordination, and tool usage, enhancing capabilities and enabling autonomous operations. The discussion encourages rethinking workflows instead of merely integrating AI into existing processes. The TACO framework categorizes agents based on goal complexity and planning needs, distinguishing between Singular Goal Agents, Automators, Collaborators, and Orchestrators.

While many organizations focus on taskers for quick wins, there is increasing interest in specific use cases and benchmarking against adoption rates of agent technologies. Recent advancements include standardizing reasoning models and introducing the Model Context Protocol (MCP) for agent communication. The KPMG Pulse Survey indicates a significant rise in agent pilots and everyday AI usage, suggesting a shift towards practical applications of agency in AI systems.

The conversation stresses the importance of safety and trustworthiness in AI agents, highlighting the necessity of ongoing maintenance from the outset. Companies are encouraged to build an agent-ready organization by centralizing decision-making around models and tools, which involves careful consideration of processes, governance, and infrastructure. A central organization is crucial for defining strategy, vision, and standardization in agent development, including guidelines for models, language, and platforms.

Essential roles in a Scrum team for building agents include domain experts, AI engineers, full-stack engineers, UX designers, and quality assurance professionals. Once an agent is built, a supporting Scrum team should conduct trusted verification to ensure thorough testing and evaluation. Agents should be treated as products, incorporating roles like product owners and system architects to facilitate structured development and collaboration.

The market shows fragmentation, with various companies focusing on different types of agents, necessitating a strong central organization to maintain standardization. This organization should select a few agent platforms, define acceptable guidelines, and create a strategic portfolio of agents across business lines. Recent data indicates a significant rise in organizations piloting or deploying AI agents, particularly in software development, audit, marketing, customer service, and finance.

The key takeaway is to reimagine existing processes rather than simply converting them into agentic systems. Future discussions are expected to address standardizing agent actions and communication protocols within enterprises, reflecting the rapid evolution of AI integration in organizations.