Every time companies believe they have mastered a particular aspect of deploying AI agents, a new challenge emerges. The integration of these agents into existing applications, business rules, and processes often reveals complexities that cannot be underestimated. This evolving understanding emphasizes the significance of surrounding systems and workflows in determining the overall success of AI initiatives.
IT leaders, as they implement agents into their production environments, discover that the intelligence programmed into these agents is merely one factor in whether their endeavors will succeed or falter. A recent incident at Meta exemplifies this reality. The company developed an AI agent designed for managing data-column onboarding within its recommendation systems, a task that necessitates interaction with multiple interconnected subsystems.
In practice, the AI agent encountered significant complications. It began to apply rules designated for one subsystem to others incorrectly, leading to unforeseen issues. In response, Meta’s solution involved reorganizing its approach by implementing 14 specialized agents. These agents operate under the guidance of a small, deterministic orchestrator responsible for routing tasks, sequencing operations, and ensuring reliability across workflows. Meta indicated that this restructuring dramatically improved efficiency, reducing the time needed for the preparation of new data-column configurations from three or four days to less than one hour.
This experience serves as a valuable lesson for enterprises leveraging AI agents; the efficacy of an agent is not solely reliant on its internal intelligence but also significantly influenced by the broader systems within which it operates. This insight leads to further discussions about agent integration and collaboration. A recent survey by OutSystems, conducted alongside CIO Dive and KPMG, revealed a contrasting perspective among executives. Among the 858 executives surveyed, 82% expressed confidence in their organizations’ abilities to govern, integrate, and monitor AI agents. However, a mere 26% felt adequately prepared to manage conflicts arising between diverse agents, systems, and workflows.
For companies like OutSystems, which develops low-code application platforms, the implications of this survey are profound. Tiago Azevedo, CIO, highlighted the necessity of maintaining “orchestration in code, not in agents.” This emphasizes the notion that while individual agents handle narrow, specific tasks, it is the surrounding coding that establishes available options and pathways for operation within set limitations.
Azevedo further stressed that authority must not rest with the agents themselves. Any actions taken by an agent, particularly regarding sensitive processes, should invoke identity and permission checks to ensure oversight. High-risk actions must undergo human review, with the ability to halt operations, investigate occurrences, or reverse decisions.
Likewise, Azevedo reported that OutSystems has not faced conflicts between agents in its internal operations, owing to its proactive approach in establishing systems that prevent such issues rather than waiting for conflicts to arise and subsequently addressing them.
Moreover, as companies explore AI’s capabilities, they must remain vigilant in updating the rules surrounding AI utilization. In Australia, Lendi Group illustrates this principle by building an orchestration platform named Aurora. This system, which integrates proprietary technology along with open-source and commercial components, incorporates risk and compliance requirements adaptable to evolving industry standards. According to CTO Devesh Maheshwari, while agents can gather data and interact with customers, responsibility for delivering significant advice remains with human brokers, underscoring the importance of retaining human oversight amid automation.
In contrast, Trust Bank Singapore’s journey revealed a different lesson. The bank’s automated runbooks, essential for guiding IT problem investigations, became ineffective as their associated systems evolved. Allowing AI to analyze data in an unstructured environment led to inconsistent outcomes. The solution involved a restructured approach where prescribed runbooks assist in evidence gathering, AI analyzes this data, and engineers ultimately determine necessary actions. CTO Srinivas Patil highlighted the critical nature of the runbooks in ensuring correct operations, asserting that if the guides are flawed, the potential inaccuracies of AI could exacerbate existing challenges.
While solutions vary among organizations, the convergence of lessons indicates a shared realization: systems surrounding AI need continual evolution alongside business processes. Failure to recognize dependencies may lead to further complications. For instance, Barclays is currently developing specialized agents aimed at enhancing its corporate customer due diligence processes, which are traditionally cumbersome and fraught with inter-system dependencies.
EY’s global vice chair, Julie Linn Teigland, recently underscored these complexities, noting that even minor changes, such as modifications to CRM records, can create ripple effects across various departments, including sales, commissions, and human resources. Agents may execute tasks without adequate awareness of broader workflows, leading to potentially detrimental outcomes.
The deployment of AI agents transcends mere implementation; it constitutes an ongoing dialogue between IT teams and business units. This collaborative approach is informed by real-world experiences and necessary adaptations. Interestingly, as organizations pursue AI to streamline operations, reduce costs, and enhance efficiency, they may inadvertently glean deeper insights into their processes, potentially reforming aspects previously miscomprehended.
In sum, the adoption of AI agents is not merely a technological shift but a transformative journey. By fostering communication between IT and business units and leveraging insights gained from AI deployments, companies can accelerate growth and navigate the complexities of modern technology with greater agility and effectiveness.