How Mellow's 11 AI Agents Work Together
Mellow is built around twelve AI agents, each responsible for a specific domain of HR operations. Unlike a single AI assistant that tries to answer any HR question, the multi-agent architecture means each agent is expert in its domain — trained on the specific policy types, data structures, and decision patterns relevant to its area. Together, they cover the full operational surface of an HR function without requiring the kind of general-purpose capability that tends to produce mediocre performance across the board.
The onboarding agent manages the new starter journey from contract acceptance through to ninety-day review completion. It knows which compliance steps are required for each contract type, which documents need to be signed, and when. It monitors progress against the onboarding checklist and flags to the HR team when steps are falling behind. It answers new starter questions about their role, their benefits, and how to navigate the organisation, drawing on the company's own handbook and policies rather than generic information.
The leave management agent processes leave requests, checks availability against the leave calendar, applies policy rules for each leave type, and updates the record automatically. It handles complex scenarios — part-day requests, requests that span public holidays, requests that trigger a policy check because of the employee's current sick leave pattern — without requiring manual HR intervention for each case. For organisations with hundreds of leave events per month, this removes the administrative layer entirely.
The payroll query agent answers the questions that employees most commonly ask about their payslip: why is my take-home different this month, how is my deduction calculated, what does this line item mean. It draws on the specific payroll data for that employee, explains the calculations in plain language, and escalates to the payroll administrator when the question requires human review of a potential error. The result is a reduction in payroll query volume that typically reaches seventy percent within three months.
The compliance monitoring agent tracks the organisation's compliance posture across the HR function: contract compliance, right-to-work documentation, probationary review completion, performance review completion rates, and any other compliance requirements the HR team has configured. It surfaces gaps before they become audit findings and prompts the responsible manager or HR team member with a specific action and a deadline. For organisations operating across multiple countries, the compliance agent monitors jurisdiction-specific requirements separately for each employee population.
The performance support agent provides managers with guidance on performance conversations, generates draft feedback based on the manager's notes, and prompts for the check-in conversations that the performance framework requires. It does not replace the manager in performance conversations — it makes the manager better prepared, more consistent, and less likely to miss the conversations that drive performance and retention.
The remaining agents cover document generation, policy question-answering, HR case management, recruitment support, analytics summarisation, and employee self-service. Each operates within its defined scope, with clear handoff protocols to human HR team members when a situation exceeds the agent's capability. The twelve agents communicate with each other so that a case that begins in one domain — an onboarding question that becomes a policy clarification that requires a document — is handled coherently rather than requiring the employee to start a new conversation.
The architecture is important. A single AI assistant trying to do all of this would be slower, less accurate, and less auditable than eleven specialists working in concert. The Mellow approach is the same one that makes specialist teams more effective than generalists in complex domains: deep expertise in a defined area, with coordination across the team.