The first wave of AI agents for HR looked like a single chatbot that could answer anything. The problem is that HR is not one thing. It is payroll, compliance, legal, documents, benefits, recruitment, and employee relations — each with its own rules, risks, and failure costs. A general-purpose agent either gives shallow answers or says the wrong thing in the wrong context.

A better model is a hub that routes work to specialist agents.

Why one agent fails

Most HR questions are not general. "Can I change my tax withholding?" is a payroll question. "What counts as a recordable incident?" is compliance. "Draft a termination letter" is legal and document work. Each answer depends on different data, different guardrails, and different accountability.

When one agent tries to handle all of them, it has to keep every domain loaded at once. That is a recipe for confusion. It may mix up provincial rules. It may apply document formatting to a policy question. It may expose private information because it cannot tell which context is safe.

The hub-and-spoke idea

Instead of one agent that knows a little about everything, you use a hub that knows only routing and a set of specialists that each know one domain deeply. The hub reads the incoming request, decides which specialist owns it, and hands off the work. The specialist operates inside its own guardrails and data scope. The hub then assembles the result into something the user can act on.

This is the same logic organizations already use with people. You do not ask your payroll clerk to draft a termination letter, and you do not ask legal to process a vacation request. Domain ownership exists for a reason. AI agents should respect the same boundaries.

What makes it hard

The routing has to be precise. A payroll question should not leak PII into the document specialist. A compliance check should not be answered by a general knowledge model. The hub has to understand enough about the request to place it correctly, but it should not need to understand the answer.

This separation of concerns is what keeps the system safe. The hub manages context. The specialists manage expertise.

A practical test

If you are evaluating an AI tool for HR, ask: When this system is wrong, which part of HR does it affect, and who checks it? If the answer is "everything, somehow," the tool is too broad. If the answer is "payroll, and the payroll specialist reviews it," the boundaries are right.

Frameworks worth knowing

• Separation of concerns — splitting a system into parts that each handle one responsibility well.
• Least privilege access — giving each component only the data it needs.
• Hub-and-spoke architecture — a central coordinator with connected specialists, common in logistics and aviation.

AI agents in HR will not succeed because they know everything. They will succeed because they know where to stop and hand off.