How AI Auto-Tagging Works in SharePoint, and Where It Needs Help

How Copilot in SharePoint (formerly Knowledge Agent) auto-tags your content with metadata, what it does, what it costs, and what still needs a human

If you asked most people what slows down finding a document in SharePoint, “metadata” is not the word they would use. But it is almost always the actual answer. A contract with no client name attached. A policy with no review date. A hundred marketing assets named “Final_v3.” None of that is a technology failure. It is a tagging failure, repeated across millions of files, in almost every organization we work with.

Microsoft’s answer to this problem is now built directly into SharePoint. What launched in preview as Knowledge Agent has, at general availability, been renamed Copilot in SharePoint, and its core promise is straightforward: AI suggests the tags, so your team does not have to do it by hand.

This article explains how that actually works, what it costs, and where a human still needs to be involved, whether you are in IT, HR, Legal, Communications, or any other role that owns content in SharePoint.

Why Metadata Matters More Than Ever

Metadata is, in plain terms, data about data. In a SharePoint document library, it is the structured information attached to a file beyond its name and content: who the client is, what category a contract falls into, when a policy needs review, which product a marketing asset belongs to. Populated metadata improves discoverability, supports workflows, and enables automation. This has been true for years, in scenarios like product catalogs, marketing assets, contracts, and legal documents.

What changed at Ignite 2025 is that metadata stopped being a nice-to-have for search and became a direct input into how Copilot reasons over your content. Copilot does not just retrieve files. It reasons over metadata, intranet pages, the full content of Office files, and even encrypted files with sensitivity labels. Large language models are creative but not always consistent, so metadata provides the guardrails for accurate, repeatable answers. When Copilot is grounded in a library with populated metadata, its answers are measurably more accurate and reliable.

This is the direct line from our SharePoint AI Control Plane article: metadata is no longer optional. It is the price of admission for getting good answers out of Copilot. The problem has always been that populating metadata by hand, across thousands of existing files, is tedious enough that most organizations never finish the job. That is exactly the gap Copilot in SharePoint is built to close.

Meet Copilot in SharePoint, Formerly Knowledge Agent

Knowledge Agent launched in public preview in September 2025, with more than 1,800 organizations signing up in the first two months. At general availability, Microsoft renamed it Copilot in SharePoint, though some underlying technical references still carry the old name during the transition.

At its core, it is a content curator built directly into SharePoint. It enriches, organizes, and structures your content so that Copilot and other agents can reason over it accurately, not just retrieve it. You will find it through a floating button that appears across SharePoint surfaces, wherever you are working: browsing a document library, editing a page, or managing a site.

The menu it opens is context aware. What you see depends on where you are and what role you have:

• As a site owner, you will see options like “Improve this site,” which help fix broken links, retire outdated pages, and fill content gaps based on how people actually search
• As a library owner or editor, you will see “Organize this library,” “Set up rules,” and “Create new views,” which cover autofilling metadata, classifying documents, generating workflows, and building views that make content easier to navigate
• As a viewer of a site or library, you will see “Ask a question,” which opens a chat grounded in the content you are currently looking at

The same button also gives you summarization, document comparison, audio overviews, and FAQ generation, all triggered through natural language rather than menus and settings screens.

How Auto-Tagging Actually Works

The mechanic behind auto-tagging is simple to describe, even if the underlying reasoning is not. You define the tag categories, in SharePoint terms, the columns. The AI suggests the values for each document, based on the actual content of the file.

A practical example from Microsoft’s own rollout: a library of resume files, previously just a folder of unstructured PDFs and Word documents, becomes an organized, AI-ready database. The agent reads each resume and suggests values for columns like years of experience, primary skill set, or most recent role, turning what was a pile of files into something you can filter, sort, and query.

This unlocks a few things that used to require manual work or a developer:

• AI-generated views. Ask for “policies expiring in 2026” or “contracts grouped by client,” and the view builds itself from the metadata that is now populated
• Natural language rules. Describe what you need, for example “email me when invoices over $500 are added,” and the agent builds the workflow. No Power Automate expertise required
• Grounded answers in context. Because the metadata is now populated, anyone can ask a question about the library and get an answer that reasons over accurate, structured data rather than guessing from file names

None of this requires you to tag anything by hand first. You define what a “client” column or a “review date” column should mean. The AI does the repetitive work of applying it, document by document, at a speed no team could match manually.

What Roles Actually Use This, Not Just IT

This is not an IT tool that happens to be useful elsewhere. Microsoft built it around real workflows across departments, and the roles that benefit most are rarely the ones sitting in IT.

HR can tag policies with review dates and categories, highlight outdated content automatically, and build views that support audits and reduce compliance risk, without a spreadsheet tracking it all separately.

Legal can classify documents, extract key clauses, and build rules that automatically route flagged files for review, cutting down the manual triage that eats into billable time.

Communications and Marketing can publish structured, branded pages in minutes instead of hours, turning bullet points into a formatted draft with images and templates, freeing up time for messaging rather than formatting.

Operations and Facilities can organize content at scale, using smart metadata and views to manage lease applications, maintenance logs, and vendor contracts with actual clarity instead of a shared drive full of loosely named folders.

Sales can get grounded, accurate answers from product docs, contracts, and specs instantly, without pinging a subject matter expert for something that is already written down somewhere in SharePoint.

The pattern across every one of these is the same. None of it required a developer, a formal IT project, or months of planning. It required someone in the role who understood what “organized” should mean for their content, and a tool that could apply that structure at scale.

What It Costs, Licensing Explained Clearly

Licensing here has shifted a few times since preview, so it is worth being precise about where things actually stand.

The base experience is included with your Microsoft 365 Copilot license, at no extra cost. This covers the floating button, the contextual menu, chat grounded in site content, summarization, and the core Copilot in SharePoint experience. There is no separate fee to access the feature itself, and as of general availability it is on by default for licensed users rather than requiring a manual opt-in.

Autofill, the actual AI metadata extraction and column population, uses a pay-as-you-go billing model. This is a meaningful distinction. Having the capability is included in your Copilot license. Actually running the extraction across a library of a thousand documents is billed based on usage, per autofill column and per document page processed. Microsoft offered free trial consumption through the preview period, but the ongoing model is consumption based.

What this means practically: the cost of tagging a small, high-value library, say a few hundred key contracts, is going to be modest. The cost of running autofill blindly across your entire tenant, on every library, regardless of whether anyone actually needs that metadata, adds up in a way that is easy to lose track of if nobody owns the decision of where to turn it on.

Where Humans Still Need to Be in the Loop

This is the part of the story that gets left out of most product announcements, and it is the part that matters most if you have read the rest of our series.

As of today, Copilot in SharePoint has real governance gaps that Microsoft has been open about improving over time, but that exist right now, in the version most organizations are using:

• Naming conventions are not enforced. Site members can accept the AI’s suggested column names, but there is little to nothing stopping inconsistent naming across libraries. One library ends up with “Client,” another with “Customer Name,” another with “Account,” describing the same concept three different ways
• New columns land in the default content type. This may not align with any custom content types or taxonomy your organization already has in place, quietly creating a second, parallel structure alongside the one you spent time building
• The AI suggests. It does not govern. Auto-tagging is genuinely good at applying structure once that structure is defined. It is not designed to decide what your taxonomy should be, or to notice when a suggested tag conflicts with an existing standard

This connects directly to the argument we have made throughout this series. AI amplifies whatever structure already exists. Auto-tagging is a powerful accelerant for good metadata practice. Left completely unsupervised across a large tenant, it is an equally powerful accelerant for inconsistent metadata practice, just faster and harder to unwind than the manual version ever was.

None of this is a reason to avoid the feature. It is a reason to make sure someone owns the taxonomy before you turn the accelerant on.

How to Roll This Out Without Making a Mess

A practical, low-risk path to adopting this well:

• Define your taxonomy first, even loosely. Before enabling autofill anywhere, agree on the handful of columns that actually matter for your highest-value content. This does not need to be perfect. It needs to exist before the AI starts filling anything in
• Pilot on one library, not the whole tenant. Site-level opt-in exists specifically so you can test this on a single contained library, see what the AI suggests, and correct course before it touches everything
• Review the first batch of AI-suggested tags before trusting them at scale. Spend twenty minutes checking whether the suggested values actually match what a human would have chosen. This tells you far more about readiness than reading a feature list ever will
• Name an owner for the taxonomy, not just the rollout. Someone needs to be accountable for whether “Client” and “Customer Name” get reconciled into one column, and whether new columns get mapped into your existing content types rather than sitting alongside them

Done this way, auto-tagging becomes exactly what it should be: the tool that finally makes metadata realistic to maintain at scale, instead of a good idea that only ever got done for the first fifty documents.

Good Metadata Was Always the Goal. This Is What Makes It Achievable.

Metadata has always been the difference between a document library that works and one that quietly becomes a place where content goes to be lost. What changed is not the value of good metadata. It is the cost of getting there. Copilot in SharePoint takes a task that used to require hours of manual tagging, or never happened at all, and makes it something a content owner can realistically finish in an afternoon.

The organizations that get the most out of this will not be the ones who turn it on everywhere and hope for the best. They will be the ones who decide, deliberately, what good structure looks like for their most important content, then let AI do the repetitive work of applying it consistently. That is the same principle that runs through everything else we have written in this series. AI does not replace the judgment of deciding what matters. It removes the excuse for never getting around to doing it.

Not sure where to start, or whether your existing SharePoint content is ready for this?

→ Book a Consultation with Impactory

Our team can help you identify which libraries would benefit most from auto-tagging, define a taxonomy that will not conflict with what you already have, and roll this out in a way that improves your content instead of adding a second, competing structure on top of it.