Cancer Council NSW has quietly launched Genie, a custom-built AI agent in Microsoft Teams that delivers fast, cited answers drawn exclusively from the charity’s approved SharePoint libraries—showing that governed enterprise AI is no longer a pipe dream but a practical deployment serving nearly 500 employees.
What Actually Changed
Genie isn’t a generic chatbot. It was built using Microsoft’s Copilot Studio, the low-code tool for creating autonomous agents and publishing them to channels like Teams. The Sydney-headquartered charity—which focuses on cancer research, prevention, and support—says the agent now serves roughly 475 staff members who can ask natural-language questions and receive answers with direct citations back to specific SharePoint documents or pages. That means every response is verifiable and grounded in the organization’s own vetted content, not the open web.
The tight integration with SharePoint is what sets Genie apart. Instead of relying on an unmanaged search index or the broad internet, the agent only pulls from approved document libraries. For a charity handling sensitive health information, that governance is non-negotiable. When an employee asks about a grant application deadline or the latest cervical screening guidelines, Genie returns an answer—and footnotes that point straight to the source file, complete with a clickable link.
Crucially, the agent respects existing SharePoint permissions. If a staff member shouldn’t see a particular document because of their role, Genie won’t surface it, no matter how relevant the answer might be. This is not a bolt-on; it inherits the Microsoft 365 security model. According to the charity, the rollout has cut down the time employees spend hunting through intranet pages and email attachments, freeing them to focus on higher-value work.
What It Means for You
For Everyday Microsoft 365 Users
If your organization hasn’t yet rolled out a custom Copilot agent, Genie gives you a glimpse of what’s coming. Imagine a co-worker who has read every policy document, grant guideline, and HR form—and can summarize them in seconds with citations. That’s Genie. It understands internal jargon, respects departmental silos, and never hallucinates sources because every claim is tethered to a real document. For workers who dread the labyrinth of SharePoint folders, this is a game-changer.
For IT Admins and Governance Teams
The real lesson here is about control. Too many AI pilots fail because they either lock down everything (making the tool useless) or open the floodgates (creating compliance nightmares). Cancer Council NSW’s approach demonstrates a middle path: curate a set of approved SharePoint sites, rely on M365’s native permissions, and let the agent do the rest. Before you build your own version, you’ll need to:
- Audit your SharePoint environment to identify authoritative sources.
- Define which libraries, pages, or metadata tags are “agent-ready.”
- Test relentlessly: does the agent ever cite obsolete documents or draft versions? How does it handle contradictory information?
- Plan for lifecycle: what happens when a document is archived? Does the agent’s knowledge base automatically refresh?
For Developers and Power Users
Copilot Studio lowers the barrier dramatically. You don’t need to train a custom model or write complex retrieval pipelines. The tool provides a graphical interface to define topics, connect to data sources (SharePoint, OneDrive, Dataverse, and later third-party systems), and publish to Teams in hours. But don’t mistake simplicity for a lack of depth. The prompt engineering—what instructions you give the agent to control tone, format, and grounding—can make or break the experience. Cancer Council NSW likely spent significant time fine-tuning the agent’s behavior, such as requiring citations for every answer and refusing to speculate when no source covers the query.
How We Got Here
Just two years ago, enterprise AI meant off-the-shelf chatbots that struggled with anything beyond FAQs. The rise of large language models changed expectations, but organizations quickly hit two walls: hallucination and governance. Microsoft’s answer has been a steady march toward “grounded” AI. In 2023, it launched Copilot for Microsoft 365, which grounds answers in users’ own emails, files, and meetings. Then came Copilot Studio in November 2023, giving organizations a way to build custom, standalone copilots with their own data connections.
Genie is a product of that trajectory. It uses the same underlying technology that powers Bing Chat Enterprise and Microsoft 365 Copilot, but with a tight scope—a single SharePoint environment rather than the entire Microsoft Graph. This aligns with the industry’s shift toward retrieval-augmented generation (RAG), where AI generates answers by searching a pre-approved knowledge base rather than relying on its internal training data. For a charity like Cancer Council NSW, where accuracy about cancer information is literally life-critical, that approach isn’t optional.
The non-profit’s move also reflects a growing trend: smaller, focused AI agents that handle specific tasks rather than a single monolithic assistant. Microsoft calls this the “copilot stack,” and it’s betting that every organization will eventually have a portfolio of custom agents—some for HR, some for IT support, some for legal compliance. Genie is an early example of that vision in the field, not just in a keynote demo.
What to Do Now
If Genie’s story inspires you to explore a governed Copilot agent in your own workplace, here’s a practical roadmap:
-
Inventory your content
Not every SharePoint site is ready for AI. Scrub for out-of-date pages, inconsistent metadata, and redundant documents. The cleaner your sources, the more reliable the agent’s answers. -
Define governance guardrails
Decide which sites will feed the agent and how you’ll handle sensitive data. Remember: the agent inherits SharePoint permissions, so if your permissions are a mess, your agent will be too. -
Start with a narrow use case
Cancer Council NSW didn’t try to answer every question across the entire charity. Pick a department or a single knowledge base—like new employee onboarding or IT self-service—and prove value there before expanding. -
Leverage Copilot Studio’s testing tools
The studio includes a test pane where you can simulate queries and inspect the cited sources. Use it aggressively to catch misfires. Involve real users early; they’ll ask questions you’d never anticipate. -
Train your people
A genie in a bottle still needs a human who knows what to wish for. Show employees how to craft effective prompts and, critically, how to verify sources when an answer matters most. Build a feedback loop so the agent improves over time. -
Monitor and iterate
Copilot Studio’s analytics dashboard can show you common queries, fallback rates (when the agent doesn’t know), and user satisfaction. Use that data to curate new content or refine instructions monthly.
What’s Next
Genie is unlikely to remain a solitary agent for long. Microsoft’s roadmap for Copilot includes autonomous agents that can not only retrieve information but also take action—triggering a workflow, updating a record, or scheduling a meeting. A future version of Genie might, for example, automatically create a draft grant proposal based on a new funding opportunity that appears in a SharePoint monitored folder.
But the immediate challenge for most organizations isn’t action; it’s trust. The world barely trusts ChatGPT to summarize a Wikipedia article. Getting employees to rely on an AI for company policy requires a level of transparency that Cancer Council NSW appears to have achieved with its citation-first design. As more case studies like this emerge, the focus will shift from “Can AI do this?” to “How do we do it safely?”—and that’s a question every IT leader should be prepared to answer.
The quiet lesson of Genie is that governed AI doesn’t have to be painful. Start small, lean on Microsoft’s built-in tooling, and let the citations speak for themselves. Because when an agent can point to the exact page and paragraph where an answer lives, it’s hard to argue with the result.