When the Singapore National Trades Union Congress (SNTUC) first tested Microsoft 365 Copilot with 50 employees in January 2025, it was a cautious toe in the water. By July 2026, the organisation had not only extended Copilot to all 1,300 staff but had also built a prototype agent that helps industrial relations officers pull labour-law information from vetted government and union sources—offering a practical blueprint for governed AI in high-stakes work, as detailed in a Microsoft Source feature.

From a pilot to organization-wide AI, with agents built for specific roles

SNTUC’s Copilot journey picked up pace after that initial trial. A year later, the deployment covered every employee, making AI assistance a standard part of the toolkit. But the headline isn’t the seat count; it’s how the organization moved from generic chat to role-specific automation.

Chief Transformation and Technology Officer Siow Shong Seng set up “Copilot Cowork” agents that run on their own schedule. Each morning, they compile updates on technology trends, the labour movement’s social media mentions, and economic and jobs news. Siow no longer types the same prompts into Copilot every day; the agents do it for him, delivering a personalised briefing without human intervention. It’s a modest time-saver that demonstrates the value of moving beyond on-demand prompting.

The more innovative effort is an internal agent built to support industrial relations officers (IROs)—the staff who represent workers in employment disputes, review contracts, and negotiate with employers. Siow created a prototype that draws information from a defined list of sources: SNTUC’s own documents, Singapore’s Ministry of Manpower, and other government agencies. An officer can query the agent about labour laws, guidelines, or procedures and receive answers grounded in approved material.

Crucially, Siow describes this agent as a “template” to give his team ideas, not a production system. That framing is deliberate. In fields where accuracy and professional judgment are paramount, an AI agent must be treated as an assistant that speeds information retrieval, not a decision-maker. The prototype’s value lies in showing what’s possible while keeping the human firmly in the loop.

What this means for IT and business leaders

For Microsoft 365 administrators and technology decision-makers, SNTUC’s experience offers three clear takeaways.

Agent value comes from curation, not just activation. Rolling out Copilot licenses across an organisation doesn’t magically boost productivity. The real gains emerge when teams identify repeatable, information-heavy tasks and configure agents that automate them. Siow’s daily briefings and the IRO prototype are both examples of workflows that previously required manual querying and now run autonomously.

Grounding is non-negotiable. The IRO agent succeeds because its knowledge base is tightly controlled. In labour disputes, quoting an incorrect regulation could have serious consequences. By limiting retrieval to official SNTUC and government sources, the organisation ensures that the information presented is authoritative and verifiable. Any business considering AI agents for legal, HR, or compliance functions should adopt a similar source-whitelisting approach.

Treat initial agents as governed prototypes. Siow’s “template” language is smart. Rather than releasing an untested agent into a sensitive workflow, SNTUC is using it as a learning tool. IT leaders should encourage small, iterative experiments with clear guardrails: define what the agent can access, who can use it, and how its output will be reviewed before it becomes part of an official process. This reduces risk and builds institutional knowledge about what works.

For business units dealing with regulated content, the SNTUC model is a safe starting point. An AI assistant that finds relevant contract clauses or policy paragraphs can save hours, but the final decision—whether to proceed with a dispute, amend a contract, or advise a member—must rest with a trained professional. The agent is a research accelerator, not an autopilot.

How Singapore’s labour movement became an AI early adopter

SNTUC represents more than 1.4 million workers in a tripartite system that brings unions, employers, and the government together to shape labour policy. In such a collaborative environment, an embrace of AI sends a strong signal: the organisation urging workers to become AI-ready is using the tools itself.

Singapore has been pushing hard on AI adoption through initiatives like AI-Ready SG, which focuses on skills training and providing access to AI tools. SNTUC’s internal rollout aligns with that national agenda but goes further. It makes AI adoption tangible—not a theoretical upskilling exercise, but a real deployment that touches the daily work of leadership and staff.

The 50-person pilot in early 2025 was a logical first step. Siow, as an early user, noted that Copilot has evolved significantly since then. Initially, he found himself copying answers from Copilot’s chat into emails or slides. Now, the AI integrates directly with Outlook and PowerPoint, eliminating that manual hand-off. It’s a reminder that the technology itself is maturing, and early piloters often have a bumpier experience than later adopters.

What your organization can do now

Based on the SNTUC playbook, here are practical steps for any enterprise looking to move from Copilot licenses to meaningful AI agents:

  1. Identify repetitive, information-gathering tasks. Look for processes where staff spend significant time searching internal databases, external regulations, or market intelligence. These are prime candidates for automation via agents.
  2. Start with a limited pilot. Before a broad rollout, give Copilot to a cross-functional team that includes both tech enthusiasts and sceptics. Their feedback will surface real-world friction points. SNTUC started with 50 users—a manageable number for close monitoring.
  3. Build agents with curated source lists. Decide which internal and external data sources are authoritative for each use case. In the IRO example, that meant specific government websites and union documents. Avoid giving agents access to the open web for tasks that require precise, verifiable information.
  4. Treat the first version as a prototype. Label it a “template” and involve domain experts from the beginning. Let them test the agent against real scenarios and document its limitations. Use that feedback to tighten grounding, refine prompts, and add missing sources.
  5. Define human-in-the-loop rules. Specify when an agent’s output must be reviewed by a subject-matter expert. For contract review or dispute advice, that might mean every instance. For daily news summaries, less scrutiny is needed.
  6. Leverage cross-app integration. Encourage users to exploit Copilot’s ability to work across Outlook, PowerPoint, and other Microsoft 365 apps. The less copy-paste, the smoother the workflow—and the higher the perceived value.

What comes next

SNTUC plans to move from prototype to production for the IRO agent, but no timeline has been shared. That next phase will be critical: scaling a governed AI tool while maintaining source fidelity and human oversight is a challenge many enterprises will face. Meanwhile, Microsoft continues to invest in agent-authoring capabilities and enterprise controls, which could make it easier for organizations to build and deploy agents with built-in compliance features.

For Windows and Microsoft 365 shops, the SNTUC case is a reminder that the most instructive AI implementations are often the quiet, methodical ones that prioritize governance over flashiness. The agent that helps a labour officer find a regulation in seconds may not make headlines, but it’s exactly the kind of practical, responsible AI adoption that turns a corporate initiative into a daily productivity tool.