Enterprises Deploying Large-Scale Agentic AI Systems
Companies are increasingly deploying agentic AI systems composed of hundreds of specialized autonomous agents for business tasks. Genesys has launched an agentic virtual agent for enterprise customer experience. Meanwhile, consulting firm Sia is now using over 400 agents on its internal agent store to accelerate development, signaling a move toward multi-agent workflows in enterprise settings.
- The Genesys virtual agent is built on Large Action Models (LAMs) and is designed to execute multi-step actions across different systems to resolve customer issues without human intervention. It is expected to be globally available in the first quarter of the company's 2027 fiscal year (February 1 to April 30, 2026). - Sia's Agent Store, accessible at siagents.ai, evolved from a Generative AI platform launched in June 2023 to a collection of over 400 specialized agents. These agents are designed for industries like finance, energy, and healthcare, and fulfill roles such as analyst, strategist, and auditor. - Evaluating agentic AI systems requires a shift from traditional metrics to assessing behaviors like multi-step reasoning, tool use, and error recovery. Benchmarks like AgentBench and ToolBench are emerging to provide standardized evaluation across different environments, including operating systems and databases. - Agentic AI creates new needs for data labeling, moving beyond simple annotation to include validating the sequence of actions, tool selection, and error handling of AI agents. High-quality, human-labeled data is crucial for fine-tuning these systems, especially for domain-specific and edge-case scenarios where synthetic data may fall short. - Reinforcement Learning from Human Feedback (RLHF) is a key technique for training models to be helpful and harmless, but it can be expensive. Constitutional AI is an alternative approach where a model critiques and revises its own outputs based on a set of principles, reducing the need for extensive human labeling. - While synthetic data can be generated faster and more cost-effectively, human-labeled data remains superior for tasks requiring nuanced understanding, domain expertise, and bias mitigation. A hybrid approach, using synthetic data for scale and human annotation for critical and complex cases, is often the most effective solution. - Go-to-market strategies for AI infrastructure startups are shifting to focus on a smaller set of high-impact capabilities like buyer-intent detection and generative personalization, rather than a broad suite of tools. Success with technical buyers often requires integrating into their existing workflows and demonstrating clear ROI through targeted use cases. - AI is projected to significantly alter the workforce, with some estimates suggesting it could replace a quarter of work tasks in the US and Europe. However, reports also predict a net gain in jobs, with millions of new roles emerging in areas like AI development, data analytics, and machine learning by 2030.
