2026从嵌入式 AI 到自主智能体：下一代人力资源服务交付架构选型指南

From Embedded AI to Lead Agents:A research-backed guide for HRTransformation and HR Tech leaders Architecture choices for the next generation of HR servicedelivery Authors:André Fortange & Volker Jacobs04May2026 Thisresearch draws on Insight222’s HR Transformation Leaders Program,consultingwork,targeted practitioner interviews, and a structured evidence review. It argues that LeadAgent architecture for HR is now credible—but only in a disciplined, specialist-first form. Thedecisive choices are not about platforms. They are about what sits below the interface. ABSTRACT The paper provides a decision framework for comparing architecture options, selecting first usecases, and avoiding premature platform commitments. HR is entering a new phase of AI-enabled transformation in which isolated pilots and workflowautomations are evolving into broader agent-based service architectures. Thisresearchpaperevaluates whether a distinct Lead Agent model is now meaningfully viable for HR: a front-dooragent that receives requests (in natural language) and dynamically coordinates specialistExpert Agents across domains such as payroll, onboarding, learning, compliance, andemployee services. The evidence suggests that this model is now buildable. But the decisive constraint is notmodel sophistication alone. Success depends on disciplined sequencing, source-systemintegration, permission integrity, auditable outputs, and adoption. The practical implication isthat a Lead AgentArchitectureis not universally superior. It is one credible design option in amarket that now includes three viable architectural patterns. What the research suggests •Lead Agent architecture is now technically credible, but only as a supervisoryrouting and orchestration layer.The evidence does not support the idea of an all-knowing“HR super-agent”.•The decisive constraints are architectural and organisational, not only model-related.Permissions, knowledge quality, API maturity, workflow design, governance,and user trust determine viability.•Three architecture patterns are now visible in the market.Embedded AI, control-tower orchestration, and Lead Agent + Expert Agents each represent credible butcontext-dependent design options.•The most robust implementation path is specialist-first.Build one or two high-valueExpert Agents, validate quality and trust, then add broader orchestration.•The operating-model implication is material.HR work shifts from transactionexecution toward service design, product ownership, knowledge governance, analytics,and human-machine orchestration. How to use this paperThisarticlecombines four evidence sources: TI People’s HR Transformation Leaders Program, consulting work with large organisations, targeted practitioner interviews with HR technologyand transformation experts, and a structured review of external vendor, analyst, and researchsources. The purpose is not to produce a statistically representative market study, but to derivea practical decision framework for HR Transformation and HRIT leaders making architecturechoices in 2026–2027. 1. Why this conversation is changing now For more than a decade, HR leaders have pursued the idea of a true one-stop shop: a single,seamless front door where employees can ask any question, initiate any process, and receive areliable answer or resolved outcome. The aspiration has always been clear. The difficulty hasbeen architectural, not conceptual. In most largeorganisations, HR still operates across afragmented landscape of core HCM systems, case-management tools, knowledge repositories,andspecialist applications for payroll,learning,and recruiting,connected by unevenintegrations. Asking employees to navigate that complexity directly has always created friction. Earlier attempts to solve the problem improved the experience at the margins, but notfundamentally. Portals and intranets created a single visual entry point, yet could not resolverequests end to end. Chatbots handled basic FAQs, but quickly reached thelimits of scriptedknowledge and weak execution capability. What is changing now is that the technology is nolonger confined to answering isolated questions inside individual systems. The emergingopportunity is broader: to receive a request in natural language, understand what the employeeis actually trying to achieve, invoke the right specialist capability, and return a coherent answeror completed action, while keeping the underlying complexity invisible to the user. That possibility is now real, but it is easy to overstate. The evidence does not point to the arrivalof a single general-purpose HR brain. It points instead to the growing feasibility of a composedarchitecture. In that model, a Lead Agent plays a supervisory role: it interprets the request,manages context, and routes work to the right specialist agents, APIs, knowledge assets,workflow engines, and source systems. The main constraints are therefore no longer modelaccess or chatbot novelty. They are theless gla