Experts Advise Crawling, Then Walking, Before Running with AI Agents

Welcome to the era of multi-agent AI systems, where the potential for boosting personal and professional productivity is immense. However, integrating these advanced generative AI (Gen AI) tools into your organization is no small feat. According to a recent Deloitte report, interest in autonomous agent development is on the rise, with 26% of organizations exploring this area. Over half of executives (52%) are keen on developing agentic AI, and 45% are looking to expand into multi-agent systems. Despite their promise, these systems are not a panacea for all challenges. The report highlights that agentic AI can significantly enhance the creation of sustainable business value by autonomously meeting objectives with minimal human intervention. However, it also notes that the hurdles faced by Gen AI—such as regulatory uncertainty, risk management, data deficiencies, and workforce issues—are magnified with agentic systems due to their increased complexity. Jim Rowan, head of AI at Deloitte Consulting, emphasized that unlike conventional bots that primarily react to inputs, agentic AI can plan, prioritize tasks, and execute complex workflows with little human oversight. Yet, he cautioned that implementing these systems can be costly, stressing the importance of robust data infrastructure, including scalable cloud platforms, advanced data analytics tools, and strong cybersecurity measures. For organizations looking to adopt AI agents, starting simple is key. Benjamin Lee, a professor at the University of Pennsylvania, suggests that companies already using generative AI for simple tasks are well-positioned to leverage agentic AI. These organizations have employees who are comfortable breaking down complex tasks into simpler ones for AI, thereby already experiencing productivity gains. Rowan recommends a phased approach—starting with a pilot program to test multi-agent systems in a controlled environment. He likens the current state of AI to that of a toddler, with agentic AI being more advanced, akin to a tween—functional and capable of executing specific functions. To further integrate AI agents, organizations should encourage the use of generative AI for simple tasks and develop strategies for breaking down complex tasks into manageable parts. This approach will make the productivity gains from intelligent agents transparent, understandable, and trusted. Rowan also advocates for the use of smaller language models over the larger ones that have dominated the Gen AI landscape. These smaller models can be more effective across various roles, from supply chain management to software development and financial analysis. Lee agrees, noting that intelligent agents can break down complex tasks into simpler ones and use specialized models to combine results into a coherent response. Quality data is crucial for the effective functioning of AI agents. Inaccurate, incomplete, or inconsistent data can lead to unreliable outputs and actions, posing both adoption and risk challenges. Therefore, investing in robust data management and knowledge modeling is essential. Workforce upskilling is another critical area, with Rowan emphasizing the need for training in both technical skills and the ability to collaborate with AI agents. A well-prepared workforce is vital to fully realizing the potential of AI agents. Continuous monitoring and improvement of AI agent performance are also necessary. This involves collecting and analyzing performance data, identifying areas for enhancement, and making adjustments to optimize performance. From a policy perspective, companies must establish clear guidelines on the use of agentic AI. Ben Sapp, global practice lead of intelligence at Digital.ai, points out the importance of determining who can use agentic AI, its permissions to interact with other agents, and the hierarchy for decision-making when systems interact or conflict. Sapp provided an example from a financial services company that uses an AI model to predict change failures. Based on the probability of failure, a human can decide whether to review the change more deeply or approve it. With agentic AI, if the failure probability is below 1%, the system can automatically approve the change, eliminating the need for human intervention and streamlining the process. In summary, while agentic AI holds great promise for enhancing productivity and creating sustainable business value, its successful implementation requires careful consideration of technical, data, workforce, and policy challenges.

Related article

Comments (8)