Building an HR policy agent in Copilot Studio: what actually works

The HR inbox in most Australian mid-market organisations is a queue of questions that have already been answered, in the enterprise agreement, the leave policy, the code of conduct, the onboarding guide. A business partner spends two hours a day pointing staff to documents that have been on the intranet for three years. The agent isn't replacing the HR team; it's giving them back that time.

Copilot Studio is the right surface for this agent. The grounding data is already in SharePoint. The agent publishes into Microsoft Teams where staff already work. And the policy library is stable enough that a well-scoped agent produces high-confidence answers without requiring an engineering team to maintain it. The following is what Frontrow has learned building this agent in production across Australian clients.

What the agent actually does

The scope that earns its seat in production: leave entitlements and accrual rules; remote work and flexible working policy; onboarding and offboarding checklist guidance; enterprise agreement summaries; performance review cycles and process; code of conduct basics; and the referral path for issues that require a business partner's judgement. That is the bounded surface. The agent answers the bounded questions confidently and hands off cleanly when the question is outside scope.

What the agent does not do: interpret grey-area employment law, provide advice on individual performance management situations, or resolve grievances. Those stay with the HR team, and the agent should say so explicitly when those queries come in.

Grounding: which SharePoint sites and how

The grounding library for an HR agent is typically three to five SharePoint sites. The HR Hub site is the primary source, the policy library, enterprise agreement documents, onboarding guides, and any role-specific supplements. A secondary source is the People & Culture site if one exists separately. A tertiary source is the Org-wide Communications site for headcount, org charts and structure documents.

The scoping decision that makes the biggest difference in production: ground the agent on specific libraries, not whole sites. A site with HR sub-folders, project files, marketing collateral and old contracts will produce confused and inconsistent answers. Tell Copilot Studio to ground on HR Policies, Enterprise Agreements, and Onboarding as three named libraries, and nothing else. Every document in those libraries should be reviewed for accuracy before the agent goes live. Outdated policies in the grounding surface as live answers.

• SharePoint site: HR Hub, libraries: Policies, Enterprise Agreements, Onboarding Guides
• SharePoint site: People & Culture, libraries: Benefits, Working Arrangements
• Exclude: Project files, historical documents, draft folders, attachments from 2019
• Document review: every source document tagged as current, owner assigned, review date set

Guardrails: sensitivity labels, role-based filters, refusal patterns

HR data sits at the intersection of the most sensitive categories in Purview's classification taxonomy. Before the agent reads anything, the source libraries need sensitivity labels applied. The working model: policies and EA documents at Internal, individual role-specific supplements at Confidential if they contain remuneration bands, and any documents containing named staff at Highly Confidential with encryption enforced. Copilot Studio respects Purview sensitivity labels at query time, if the label says the document is out of scope for this user, the agent will not surface it.

Role-based filtering matters here in a specific way. The agent should be accessible to all staff for general policy queries, but a query from a manager about another employee's leave balance should be refused, not answered. That is a Workday or Chris21 query, not a SharePoint document query. The refusal pattern needs to be explicit in the system prompt: 'If asked about an individual's personal leave balance, employment terms, performance status or compensation, decline and refer to HR direct. Never speculate about individual entitlements from policy documents.'

The refusal pattern library should be built out before launch and tested with a set of at least 20 queries that should fail. Common failure modes: the agent infers an individual's entitlement from a policy document; it gives a confident wrong answer on an enterprise agreement clause with ambiguous wording; it fails to distinguish between a policy that has been superseded and the current version. All three happen in early production and all three are fixed by tightening the system prompt and the source library review.

What Frontrow has shipped: professional services, eastern Australia

A professional services firm of around 300 staff across three offices had two HR business partners fielding approximately 80 policy queries per week via email and Teams messages. The queries were almost uniformly policy-level, leave accrual, public holiday treatment, onboarding timelines, performance cycle dates. The HR team's frustration was not volume, it was that the answers existed in documents they had maintained carefully for years and nobody was reading them.

Frontrow built an HR policy agent grounded on four SharePoint libraries, with a system prompt scoped to the 12 policy categories the team identified as the highest-frequency queries. The agent was piloted with the Sydney office for three weeks before broader rollout. At the six-week mark, the HR team reported a 60% reduction in routine policy queries via email, with the remaining queries skewing toward the grey-area and individual-circumstance questions that genuinely required a business partner. The business partners now spend less time on repeatable answers and more on the work that requires them.

The grounding maintenance problem

The agent is only as current as its source library. The single most common way an HR agent degrades after launch is policy documents that are updated on SharePoint without the agent's scope being reviewed. A leave policy amended by an enterprise bargaining outcome, a new remote work framework introduced mid-year, a benefits change, if the document in the grounding library changes, the agent reflects the change. If the document is not updated, the agent gives the old answer with confidence.

The operating model fix is a quarterly review of the source library by the HR team, with a clear owner for each document category. Frontrow recommends tagging each document in SharePoint with a 'Last reviewed for agent' column and setting a Planner reminder for each category owner. It is low overhead, and it prevents the agent from becoming a liability six months after launch.