Work smarter not harder: Build an AI agent in 5 minutes [Hosted by IBM]

Unleashing Enterprise AI: IBM’s Strategy to Conquer Agent Sprawl and Drive Real ROI

Mr. Sabtain Khan, Product Manager at IBM, addressed Web Summit Lisbon 2025, outlining IBM’s focus on AI agents for large enterprises like banks and airlines. He detailed the evolution from basic predictive models to sophisticated, contextual AI assistants, emphasizing the need to handle complex, multi-intent user queries effectively, moving beyond the limitations of simple chatbots. The rapid proliferation of AI agents from various vendors (ServiceNow, Salesforce, Microsoft) and open-source solutions has created a significant “agent sprawl” problem, leading to disconnected user experiences and hindering overall efficiency and satisfaction within large organizations.

Building an AI agent is merely 20% of the effort, Mr. Khan explained. The remaining 80% involves extensive testing, rigorous evaluation, and complex integration with existing enterprise backend systems, which often entail navigating VPNs, private networks, and strict regulatory compliance. Continuous monitoring and optimization based on real-world feedback are paramount for long-term success.

Enterprises often expect rapid ROI, attempting to implement AI across all use cases simultaneously. Mr. Khan advised a design-first approach: define clear goals and Key Performance Indicators (KPIs), then prioritize high-value, lower-effort use cases for iterative improvement. Establishing milestones and utilizing monitoring dashboards are crucial to demonstrate tangible return on investment, ensuring solutions address real problems.

IBM’s solution to agent sprawl is an orchestrator agent designed to unify disparate AI systems. This central agent intelligently routes user requests to the most appropriate underlying agent (e.g., HR queries to SAP, IT issues to ServiceNow) while seamlessly maintaining conversational context. This creates a single, cohesive interface, significantly enhancing the employee and customer experience.

IBM promotes an open approach, allowing integration of third-party and open-source agents, and choice of AI models (e.g., Gemini, OpenAI). This prevents vendor lock-in and leverages existing investments. The framework supports open protocols like A2A (Agent-to-Agent) connections, enabling comprehensive AgentOps monitoring to track performance, identify failures, and manage latency across the entire AI ecosystem.

Mr. Khan stressed that not all tasks benefit from a fully agentic AI solution, especially critical, deterministic processes like loan applications, where hallucination is unacceptable. He advocated for a hybrid approach, strategically injecting AI (e.g., for data mapping) into existing, pre-built automations. This spectrum of implementation ensures reliability and aligns with enterprise comfort levels.

A critical lesson is the necessity of a design-first development lifecycle. Enterprises must define clear goals and Key Performance Indicators (KPIs) before building, ensuring AI solutions address real problems and deliver measurable value. The process requires robust testing with customer data, deployment in separate environments, and continuous monitoring and optimization to ensure agents remain effective and adaptable.