AI Agents for HR: What They Can and Can't Do
AI agents are a different category of technology from AI assistants. An assistant answers a question. An agent takes a task, executes a sequence of actions to complete it, and reports back. For HR, this distinction matters: an AI assistant that answers "what is the leave policy?" is useful but limited. An AI agent that takes a new starter record, checks the compliance requirements for their role, sends the required documentation, schedules the induction calendar, and flags any missing information — and does all of this without being prompted step by step — is a different class of tool.
What AI agents can currently do well in an HR context: process structured requests end-to-end (onboarding checklists, leave applications, document generation), answer complex policy questions by synthesising information from multiple sources, flag anomalies in payroll and attendance data, draft standard correspondence, schedule and remind, and monitor compliance deadlines across a large employee population. These are real capabilities that produce real operational value.
What AI agents genuinely cannot do: exercise moral judgement in complex disciplinary situations, build trust with an employee who is in crisis, assess cultural fit in a hiring decision, manage sensitive interpersonal dynamics, or take accountability for a decision that affects someone's employment. These limitations are not temporary gaps that the next model version will close — they reflect a fundamental difference between pattern recognition and situated judgement. An AI agent can identify that a disciplinary situation has occurred; it cannot decide whether the employee's explanation is credible.
The value of AI agents in HR is therefore in the range of tasks that are structured enough to be automated but complex enough that a human would otherwise spend significant time on them. A single AI agent handling leave requests across a workforce of three hundred employees — answering queries, processing applications, flagging policy conflicts, updating records — frees the equivalent of a full HR administrator's time. That is not a marginal efficiency gain.
Designing AI agent workflows for HR requires careful thought about where the human handoff should occur. A workflow that routes all leave queries to an AI agent, but escalates to a human when the query involves a medical situation or a family crisis, is appropriately designed. A workflow that routes all leave queries to AI without exception, including sensitive medical requests, is not. The handoff criteria should be explicit, reviewed regularly, and calibrated against the cases that have been escalated to understand whether they were handled correctly.
Mellow runs twelve AI agents across the core HR function: an onboarding agent, a leave agent, a compliance agent, a payroll query agent, a performance support agent, a document generation agent, a policy question agent, and more. Each agent has a defined scope, defined handoff criteria, and a full audit trail of every action taken. HR teams can see exactly what the agents have done, override any decision, and reconfigure scope as the organisation's needs change.
The practical starting point for most organisations is one or two agents in the highest-volume query areas. Leave and policy questions are typically the most common HR queries and the ones where AI handles the volume most effectively. Starting there, measuring the impact, and expanding the agent scope based on evidence is a more sustainable approach than attempting to automate the entire HR function simultaneously.