Agentic AI for HR: Copilot Studio Agents Grounded in Your SharePoint HR Policies

Short version: Generative AI helped HR write faster. Agentic AI lets HR delegate: an agent that answers policy questions from your own documents, walks a new hire through onboarding, checks a leave request against the rules and drafts a job description in the house style — then hands anything with judgement in it to a person. The tool that makes this practical for a Microsoft 365 organisation is Copilot Studio, and the ingredient that makes it trustworthy is grounding the agent on the HR policies you already keep on SharePoint, so every answer is cited and permission-trimmed. This guide covers the five HR agents worth building first, a build walkthrough, the readiness work most teams skip, and the legal and ethical guardrails Singapore HR must keep. The hands-on version is the WSQ Agentic AI for HR course (TGS-2024045795), funded up to 70%.

From generative to agentic: what actually changes for HR

Two years ago the HR use case for generative AI was a text box: paste a policy, ask for a summary, tidy the tone of a difficult email. Useful, but the human still did every step. An agent is different in three ways. It has a goal rather than a single prompt, it can use tools (look something up, create a record, start a workflow), and it can be grounded on a specific body of knowledge so it answers from your policies rather than from the internet’s idea of a policy. Microsoft’s overview of agents for Microsoft 365 Copilot describes them as Copilot customised for a specific scenario, with its own instructions, knowledge and actions, which is exactly the shape HR needs.

The right-hand column is not science fiction. Every row is a Copilot Studio agent with a SharePoint knowledge source and, at most, one Power Automate flow. We covered the broader “why now” in an earlier post on agentic AI for HR in Singapore; this article is about how to build it.

Why SharePoint grounding solves HR’s biggest fear

Ask any HR director what stops them deploying an AI assistant and the answer is the same: it will make up a policy. A confidently wrong answer about notice periods or medical leave is worse than no answer. Grounding is the fix. When you add a SharePoint site or library as a knowledge source in Copilot Studio, the agent’s generative answers are produced from the content of those documents, with citations back to the file. If the answer is not in the documents, a well-instructed agent says so and offers to connect the person to HR instead of improvising.

Two properties make this safe for HR specifically:

• Permission trimming. The agent answers with the permissions of the person asking. A staff member cannot extract the executive compensation policy through the agent if they could not open it on SharePoint. Copilot Studio’s security and governance documentation covers the authentication settings that enforce this.
• One source of truth. When the policy changes, you update the document and the agent’s answers change with it. There is no second copy inside the agent to drift out of date.

The consequence is that the first project in agentic HR is not an AI project at all. It is a SharePoint project: one HR policy library, current versions only, clear file names, superseded documents archived out of the library, and access reviewed. The SharePoint Online for End Users course exists for teams that need to fix this first. Do it before the agent, and the agent is good on day one.

The five HR agents to build first

1. Policy Q&A agent. Grounded on the employee handbook and leave, benefits, claims and code-of-conduct policies. Published in Teams. Measured by the drop in HR mailbox volume. This is the one to build first because it is the simplest and the most visible.
2. Onboarding buddy. Grounded on the onboarding guide, IT setup instructions and the org chart. Runs a checklist topic that walks the new hire through week one and flags overdue steps to the manager through a flow.
3. Leave and benefits assistant. Answers entitlement questions from policy, then collects a leave request and hands it to the existing Power Automate approval flow. The agent never approves; the manager does, with the balance and the policy clause shown.
4. Job description drafter. Grounded on role profiles, the competency framework and the JD template. Produces a draft for the hiring manager that HR then reviews for fairness and compliance.
5. L&D navigator. Grounded on the internal learning catalogue and a list of funded external courses. Maps a stated skill gap to options and starts the nomination workflow. Pairs naturally with the WSQ courses your staff can attend with funding.

Each of these is a two-day build for a first-timer and a half-day build once you have done one. The course has learners build the policy agent end to end and design the other four.

Build walkthrough: the HR policy agent

1. Prepare the library. One SharePoint document library, HR Policies, current versions only, readable by all staff.
2. Create the agent in Copilot Studio. Instructions along the lines of: “You answer employee questions about HR policies using only the HR Policies library. Always cite the policy document. If the answer is not in the documents, say so and offer to raise a ticket with HR. Never give advice on disciplinary matters, grievances or medical conditions; direct those to a named HR partner.”
3. Add the knowledge source — the HR Policies library — and enable generative answers.
4. Test with the ten most common questions from last quarter’s HR mailbox. For each wrong or vague answer, fix the document first; only tune the instructions if the document is already clear.
5. Add the escalation topic. “Talk to HR” creates a ticket in a SharePoint list through a flow and tells the employee who will respond and when.
6. Restrict what it will not touch. Add topics that recognise grievance, harassment, medical and termination questions and route them straight to a person, with no generative answer.
7. Publish to Teams for a pilot group of 30 to 50 staff, and review the analytics weekly: unanswered questions are gaps in the library.

Step six is the one HR teams thank us for later. An agent that is excellent at leave entitlements and silent on grievances is a good agent. An agent that attempts both is a liability.

The readiness work most teams skip

Topic three of the course is about preparing the workforce, and it is there because technology is rarely the reason an HR AI initiative stalls. The checklist we use:

Microsoft’s 2025 Work Trend Index describes organisations moving towards human-agent teams where managers increasingly direct agents alongside people. For HR that is not a distant trend; it is a job-redesign question for the HR team itself, and the readiness checklist is how you start answering it.

Legal, ethical and security guardrails for Singapore HR

Topic four of the course is the one that keeps the initiative defensible. The non-negotiables:

• PDPA. Employee data is personal data under the Personal Data Protection Act. Keep the agent’s knowledge to policies and procedures, not personnel files; where an agent must look up an individual’s record (a leave balance), do it through an authenticated flow that returns only that person’s data. The PDPC publishes advisory guidelines on the use of personal data in AI systems that HR should read before scoping anything that touches decisions about people.
• Fair employment. The Tripartite Alliance for Fair and Progressive Employment Practices guidelines apply to AI-assisted processes exactly as they apply to manual ones. No agent should rank, shortlist or reject people on its own; it can prepare information for a human who decides.
• Human oversight by design. Every agent answer that concerns an individual’s employment carries a path to a named person. Adverse decisions are never automated.
• Transparency. Staff know they are talking to an agent, know what it is grounded on, and know how to reach HR.
• Security. Authenticated users only, knowledge-source permissions reviewed, editing rights on the agent limited to the HR ops owner, and tenant-level data loss prevention in place.

None of these are obstacles to the five agents above. They are the reason those five were chosen: they inform and route; they do not judge.

A 90-day roadmap

Learn it hands-on, funded up to 70%

The WSQ – Agentic AI for HR course (TGS-2024045795) is a two-day programme designed for HR professionals and people leaders, not developers. Its four topics follow the arc of this article:

1. Catalysing HR with generative AI — the future of work and how AI transforms and amplifies HR functions compared with conventional practice.
2. Generative AI applications in HR — talent acquisition and onboarding, leadership and employee development, diversity, equity and inclusion, employee experience and HR policy development.
3. Preparing the workforce for generative AI — assessing organisational readiness, building AI awareness and fluency, identifying skill gaps, training and pilot testing.
4. Legal considerations and the future of generative AI — legal and ethical responsibilities, and where the tools are heading.

Learners build a working policy agent grounded on a SharePoint library during the labs and leave with a feasibility analysis for their own organisation. Assessment is a written and a practical exam.

The nett fee can be further offset with SkillsFuture Credit for eligible Singaporeans, PSEA for eligible younger learners, and SkillsFuture Enterprise Credit (SFEC) for eligible Singapore-registered companies; employer-sponsored SMEs can also claim absentee payroll. The course runs over two full days (9:30am to 6:30pm), with physical classroom and synchronous online (Zoom) options. Figures are taken from the course page at the time of writing — the live page always carries the current fee and the next available dates.

See the full AI for HR course list, or pair this course with WSQ Generative AI for Interviewing for the recruitment side and WSQ Business Process Automation with Power Automate and Copilot Studio Agents for the workflow side.

Frequently asked questions

The bottom line

HR does not need an AI that is clever. It needs one that is right, cited, permission-aware and honest about its limits. A Copilot Studio agent grounded on a clean SharePoint policy library is exactly that, and it removes the hundred repeated questions a month that keep HR from the work only people can do. Build the policy agent first, keep every decision with a human, do the readiness and compliance work in the open, and agentic AI in HR becomes a service staff prefer rather than a tool they were told to use.

Ready to build your HR agent? Register for WSQ – Agentic AI for HR — 2 days, funded up to 70%, SkillsFuture Credit claimable.