Bloody agents

Gartner predicts that over 40% of AI agent projects will be abandoned within the next year, mainly due to a lack of clear business value. Companies are rushing to deploy autonomous systems driven by fear of missing out (FOMO) without establishing measurable outcomes. These agents often follow instructions without being linked to strategic goals, resulting in what Jan van Niekerk, senior vice president of Data and AI strategy at Netsurit, calls "vanity work."

This refers to visible tasks that demonstrate an organization's AI usage without delivering significant cost savings or improvements.

South African enterprises face a similar challenge, licensing the same frontier models as competitors and using pre-built agents for well-documented processes. According to van Niekerk, enterprises typically go through four stages of AI maturity: personal AI (where tools are given to everyone without demonstrating P&L impact), workflow AI, process AI, and finally, being AI-native.

For enterprises to successfully deploy AI agents, they must first map and redesign existing processes, assigning ownership of the value created. Measuring the capacity freed up by the agents and tracking what it is redirected to is crucial for demonstrating value. Otherwise, reported metrics like "hours saved" may be meaningless without context on how the saved time is being utilized.

Nazia Pillay, managing director at SAP Southern Africa, emphasizes that measurable improvement should be evident within months when an enterprise picks a well-defined process. She suggests starting with the outcome and not the technology to demonstrate value and determine expansion opportunities. Key indicators that a process is not ready for an agent include lack of clear end-to-end description, high exception rates, undocumented handoffs, and absence of a single owner or baseline metric.

An example of successful AI agent deployment is in credit decisioning and fraud monitoring, where clear rules already exist, making banks prime candidates for agent implementation. However, enterprises must also address agent containment, defining the limits on what an agent can access, modify, or disrupt to prevent potential issues.

David Pretorius, customer advisor manager at SAS, highlights the importance of measuring agent density, similar to how agents are counted against people in infrastructure planning. By 2030, enterprises will need to consider the ratio of agents to people and the payoff rate, ensuring clear thresholds for predictable work while maintaining human control over sensitive decisions.