HR service agents are among the most common enterprise agent projects, and for good reason. Employees ask the same questions repeatedly, the answers sit in policies and records the organisation already holds, and many staff, particularly in manufacturing, retail and logistics, never sit at a desk. Meeting them on a messaging channel they already use is a genuine service improvement.
The demonstration is easy. Point a model at the leave policy, connect it to a chat window, and it will answer "how many days of family responsibility leave do I get?" convincingly within an afternoon.
The production system is different. It has to know which employee is asking before it says anything personal. It has to answer from the current policy, not last year's. It has to read the right record from the HR system, open a case when something needs to change, and recognise when a conversation should go to a person immediately. Each of those is an integration with an existing system that has its own owner, rules and failure modes.
CloudNala's view is that an HR agent's quality is decided mostly by its integrations and routing rules, and only partly by its model. Plan, staff and budget it as an integration project with a conversational front end, not as an AI project with some connectors.
Karoo's integration surface
Karoo Mills is fictional, and its systems and figures are illustrative. Karoo's HR agent serves about 6,000 employees across nine sites. Production staff mostly use WhatsApp; office staff use Teams; a small number use the HR portal. Behind the agent sit a policy library maintained by the HR policy team, an HR information system holding leave balances and employee records, a payroll system, a case-management system and the HR officers at each site.
The agent is the smallest box in that picture. Everything else either existed before the agent or is a contract between the agent and something that existed before it.
Channels: the same person on three doors
Each channel identifies people differently. Teams knows the employee through the organisation's directory. WhatsApp knows only a phone number. The HR portal knows a login. The agent must turn each of these into the same employee record, and must handle the cases where it cannot.
Phone numbers are the hard case. Numbers change, are shared within families, and are recycled by networks. Karoo's rule is that a WhatsApp number on its own is enough to answer general policy questions, but not enough to show personal information. For anything personal, the employee completes a step-up check: a one-time code sent to the number registered on the HR system, combined with something only the employee should know, such as their employee number. Once verified, the session carries that identity for a limited time.
The channel also limits what the agent can do. Messaging channels have their own rules about message templates, response windows and what may be sent unprompted. Those rules shape the design of reminders and status updates, so they belong in the requirements from the start.
Identity before anything personal
Identity is not a feature of the agent. It is a precondition for every personal answer and every action. Karoo treats it as a separate component, owned by IT security, that the agent calls and cannot bypass. When the agent asks the HR system for a leave balance, the request carries the verified employee's identity, and the HR system's own access rules decide what comes back.
That second point matters. The agent should not be the only thing standing between an employee and another employee's payslip. If the HR system enforces access by employee, a routing mistake in the agent produces an error rather than a data breach. Our MCP in plain language article makes the same argument for tool access generally.
Policy retrieval: current, cited and owned
Policy questions look like the simplest part. They are where agents most often give confident, outdated answers.
Karoo's policy library had three problems common to most organisations: superseded versions still sitting in shared folders, site-specific variations (shift allowances differ between mills and bakeries), and policies that refer to collective agreements held elsewhere. The fix was not a better model. It was a content job:
- The HR policy team marked one version of each policy as current, with an effective date, and archived the rest out of the agent's reach.
- Each policy was tagged with the sites and employee groups it applies to.
- The agent was required to cite the policy and section it relied on in every policy answer, so employees and HR officers could check it.
That work took longer than building the agent. It is also the work that determines whether employees trust the answers.
Routing: decide the kind of request first
The most important design decision in Karoo's agent was made by HR, not by engineers. Before the agent does anything else, it classifies the request.
General policy questions are answered with a citation. Requests about the employee's own records require identity verification first. Change requests (leave corrections, banking details, site transfers) open a case in the case-management system with the facts gathered, and the case system owns the process from there, as described in Part 1.
The last row is the one that matters most. Grievances, medical matters, harassment and anything suggesting distress go to a person immediately. The agent does not try to resolve them, gather details beyond what is needed to route, or summarise them in ways that could later be read as the organisation's position. HR wrote that rule, tested it with real examples, and reviews misrouted conversations every week.
POPIA as design questions
An HR agent handles employee personal information, and some of what employees say will be special personal information under POPIA, such as health details mentioned when asking about sick leave. This article is general information, not legal advice; Karoo's legal and privacy team made the determinations. The design questions they asked are useful for any organisation:
- What personal information does each integration return, and does the agent need all of it?
- Where are conversation logs stored, for how long, and who can read them?
- Does any information leave South Africa during processing, and on what basis?
- How are medical or other sensitive details handled when an employee volunteers them?
- How does an employee find out how the agent uses their information?
Our Which data can AI touch? guide offers a plain-language way to sort these questions before choosing tools.
Escalation is part of the product
When the agent hands a conversation to an HR officer, the handover should carry the verified identity, the request type, the facts gathered and the conversation so far, so the officer does not start again. Karoo measured its handover quality by one number: how often an HR officer had to ask the employee for something the agent had already been told. In the first month it was roughly one handover in three. After the handover record was redesigned, it fell to under one in ten.
That measure says more about the employee experience than any model accuracy score.
A planning checklist for an HR agent
| Workstream | Owner | Typical first deliverable |
|---|---|---|
| Channel integration | Digital or IT | Identity mapping per channel; message rules |
| Identity and step-up checks | IT security | Verification flow; session lifetime |
| Policy content | HR policy team | One current version per policy, tagged by site and group |
| HR system access | HRIS owner | Read interfaces scoped to the verified employee |
| Case integration | HR operations | Case types, required fields, status feed |
| Routing rules | HR leadership | Written rules and a test set of real examples |
| Privacy | Legal and privacy | Determinations on retention, transfer, sensitive data |
| Agent and evaluation | Delivery team | Agent, evaluation set, tracing |
Notice how few rows belong to the team building the agent. That is the point.
Next: Part 4, Capability architecture, not another AI logo landscape. All parts are on the series page.
How CloudNala can help
We help organisations scope HR and other employee-service agents as the integration projects they are: channel and identity design, policy content readiness, scoped access to HR systems, case integration, routing rules and handover design. We work provider-neutrally, so the plan holds whichever agent platform you choose.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an Agentic AI Architecture Review or write to us at consult@cloudnala.co.za