The Situation

As the organization grew, so did the volume of routine HR questions — attendance records, leave balances, parental leave entitlements, payroll details, and policy clarifications. These were handled through email, ticketing systems, or direct conversations, leaving HR business partners to spend a disproportionate share of their time on low-complexity queries rather than strategic workforce initiatives.

The same fragmentation showed up for people managers, who needed to track attendance and leave patterns across rolling six- and twelve-month periods — data that existed in the HR system but wasn't easily surfaced in a query-able form. New joiners faced the sharpest version of the problem: onboarding required up to ten hours of guided training just to learn where to find internal HR information.

The Approach

The team built the HR agent as a conversational tool embedded directly in the client's internal chat platform, backed by a Retrieval-Augmented Generation architecture and a live-data integration layer for querying HR systems in real time.

Two design choices carry the weight of the case: personal-scope, credential-based access — each employee authenticates with enterprise credentials, and the agent only returns data that person is authorized to see, enforced in application code rather than left to prompt instructions — and mandatory source citation, where every answer ends with a reference to the specific policy document, section, or data range it drew from.

We rolled the agent out in two phases rather than all at once.

Phase 1 can look things up but cannot change anything, so it answers questions, shows managers attendance and leave records, and explains company policy.

Phase 2 will let the agent act on what it finds, from booking leave to raising an issue and following it through until it is closed. Today the agent tells you things and tomorrow it will do them for you.

The Results

Every answer ends with its source, so you can see the exact policy document, section, or date range it came from. With payroll and legal rights on the line, staff needed to check an answer rather than just trust it.

Each employee sees only their own data, and that rule lives in the code rather than in the agent's instructions. The difference matters, because instructions can be talked around but code cannot. So however the question is phrased, the agent will not return a colleague's salary.

Key Takeaways

• Enforce data boundaries in code, not in the prompt. A credential-scoped access model survives adversarial input in a way that prompt-level instructions don't.
• In a domain where accuracy has payroll or compliance consequences, treat source citation as infrastructure, not a nice-to-have.
• Ship the read-only version first. Establishing trust in the agent's answers before it can take action on someone's behalf lowers the risk of the transactional phase.
• Design the first use case as one module of a shared platform, not a one-off build — it's what made extending the same agent into other departments faster later.