An AI knowledge platform that answers from approved enterprise content

The situation

Our client is a large enterprise with 28 business departments, and between them they held an enormous amount of hard-won knowledge: policies, procedures, technical references, project records, the documents people actually reach for when they need to get something right. The problem was where it all lived.

Years of content had accumulated across dozens of SharePoint sites and document libraries, each set up at a different time by a different team, with no common structure. The same document existed in three places, often in three versions, and nobody could say which was current. Search returned noise. Governance was manual, so ownership drifted, revisions went unmanaged, and expired material sat alongside approved material with nothing to tell them apart. People stopped trusting the repositories, and the knowledge that was meant to be shared quietly went back to living in inboxes and personal drives.

They were interested in AI, and specifically in giving people a way to simply ask a question and get an answer. But they were clear-eyed about the risk. An assistant that answers confidently from outdated or unapproved content is worse than no assistant at all, especially at their scale and under their compliance obligations. The real task was not to bolt an AI on top of the mess. It was to make the content worthy of it.

What we did

We treated this as a knowledge platform, not a chatbot. The whole approach rests on a simple idea: an AI is only as trustworthy as the content it is allowed to read. So we spent the effort on the content first.

• We rebuilt the information architecture. We designed a scalable structure of site collections and document libraries, then put an enterprise taxonomy and managed metadata underneath it. Content types, site columns and document sets gave every document a consistent, machine-readable classification. We optimised the search schema, defined the managed properties that mattered, and configured enterprise search so that relevance came from metadata, not guesswork.
• We built the intelligence layer on Microsoft Copilot. Once the content was properly structured and indexed across all 28 departments, we used Copilot to give people natural-language retrieval: ask a question in plain words, get a context-aware answer, with the underlying documents cited. Crucially, we grounded it. Copilot draws only on approved enterprise documentation, and the metadata we had put in place made the retrieval sharper and the results more relevant.
• We engineered the experience in SharePoint itself, using the SharePoint Framework, the Microsoft Graph API and the SharePoint REST APIs. Custom client-side web parts and extensions gave people a clean, responsive interface on modern pages, so the platform felt like part of Microsoft 365 rather than a separate tool to learn.
• We automated governance with Power Automate, which is what makes the whole thing safe to trust over time. Flows monitor document expiry, validate metadata, track ownership, and notify owners when something needs review. Revision control, approval lifecycles and version management run as workflows, with compliance reminders and an audit-ready history of every approval. This is the mechanism that keeps unapproved or expired content out of Copilot’s reach, automatically.

Security was designed in throughout. Authentication runs through Microsoft Entra ID, document ownership is role-based, and the AI inherits SharePoint’s search permissions, so people can only ever get answers from content they are entitled to see. Metadata governance, controlled approvals, version control and lifecycle management give the organisation a defensible, auditable position.

The outcome

The enterprise now has a single, governed knowledge platform spanning all 28 departments, with more than 150 GB of content indexed and searchable. The silos are gone. Instead of hunting across sites and second-guessing which version is right, people ask a question and get a trusted answer in seconds, drawn only from approved documentation.

The governance that used to eat manual effort now runs itself: expiry, ownership, revisions and approvals are handled by automated workflows, so the content stays current and compliant without someone policing it. Discoverability improved, adoption followed, and the routine burden of document management fell away.

Just as important, the organisation now has the foundation for whatever it wants to build next. Because the information architecture, metadata and governance are in place, further AI capability can be added on solid ground rather than on sand. That was always the point. Real AI adoption starts with well-governed, structured, searchable content, not with the assistant.