Copilot in SharePoint: The 20-Item Search Ceiling

Copilot search limits in SharePoint come down to one number most people building an agent never look up before they start: twenty. That is the entire knowledge source an agent can draw from, and it changes what kind of agent is actually worth building.

This covers the source ceiling, what file types actually work today versus what is still coming, and the permission behaviour that means the same agent can answer the same question differently depending entirely on who is asking.

The Twenty Item Ceiling

Microsoft states this directly: agents allow up to 20 source items as the knowledge source. Items combine across types, and the example Microsoft gives is 4 folders plus 16 files.

The number that catches people out is not the 20 itself, it is what counts as one item. A folder containing 500 documents still counts as a single source item toward the ceiling, but that also means an agent built from a handful of well-organised folders can cover far more actual content than one built from 20 individually selected files.

What Actually Reads Today

The supported list is broader than most people assume for text-based formats, and it has real gaps for anything visual. Officially supported today: DOC, DOCX, PPT, PPTX, XLSX, FLUID, LOOP, PDF, TXT, RTF, ASPX, HTM, HTML, ODT and ODP.

Images, meeting recordings, videos, and OneNote notebooks were all listed as coming soon at the time this was checked, meaning not working yet rather than working with limitations. Including one of these in an agent’s sources today does not make its content available; it simply sits there without contributing anything.

Permission Trimming: One Agent, Different Answers

Microsoft’s wording here is unambiguous: if you don’t have permission to specific content, even if this content is included in the agent’s sources, you will not see information from this content in your chat with the agent.

This is security trimming applied per user, per question, and it happens silently. The agent does not tell a user that a source was withheld because they lack access to it. It simply answers using whatever subset of the 20 sources that specific person is permitted to see.

This matters most for troubleshooting. If an agent seems to be missing something a colleague swears it answered correctly for them, the first thing to check is not the agent’s configuration. It is whether the two of you actually have the same underlying permissions to the source content.

The Rich Text Editor Is a Different Surface Entirely

Copilot’s rich text editor in SharePoint, used for drafting page content directly, operates under its own separate rules that have nothing to do with agents or their sources.

Microsoft recommends keeping the total word count in each rich text editor session under 3,000 words, and states plainly that this editor has no access to your documents or Microsoft Graph data, meaning no emails, chats or calendar content. It is a standalone drafting tool, not a window into an agent’s knowledge sources.

Licensing runs on two possible paths

Access requires either an assigned Microsoft Copilot licence, or an organisation with Pay-As-You-Go billing configured. Either path unlocks agents in SharePoint, and it is worth confirming which one applies to your account before assuming a missing feature is a rollout delay.

Building an Agent That Fits the Ceiling

Twenty items sounds restrictive until you plan around it deliberately rather than discovering the limit mid-build. A few habits make the difference between an agent that covers a department properly and one that runs out of room halfway through setup.

Start from folders, not from search results

Selecting files one at a time from search results feels precise and burns through the ceiling fast. Starting from the folder structure your team already uses, and pointing the agent at the two or three top-level folders that actually matter, usually covers more ground for far less of the budget.

Separate agents beat one overloaded agent

A single agent trying to cover an entire department’s onboarding material, policy documents and project archives at once will hit the 20 item ceiling immediately. Several narrower agents, one per topic area, each fit comfortably within the limit and tend to give more focused answers as a side effect.

Revisit the source list as content moves

A folder-based source stays current automatically as files are added inside it. An individually selected file does not update if it gets replaced by a newer version saved elsewhere, or moved to a different folder. Preferring folder sources over individual files also reduces how often the agent’s configuration needs revisiting.

What the Output Actually Is

Microsoft’s own disclaimer for this feature does not soften anything: AI-generated content may be incomplete, inaccurate, or out-of-date, and should not be relied on without independent verification. A separate line rules out an entire category of use: Copilot in SharePoint is not intended for medical, legal, financial, or other professional advice.

The same discipline that applies to any generated summary applies here. We covered the equivalent problem for Word reference documents in why Copilot ignored half your reference document, where content position and format determined what actually got used. The lesson transfers directly: know what an agent can and cannot see before trusting what it says about either.

Troubleshooting an Agent That Seems to Know Less Than It Should

Four checks cover most cases where an agent’s answers feel thinner than expected, and working through them in order rules out the cheap explanations before assuming something about the agent’s build is wrong.

The fourth check catches a specific pattern: an agent that worked well when it launched and has quietly become less useful over months. If its sources were individually selected files rather than folders, new documents created since then were never added to its knowledge, and nothing about that failure is visible from inside a normal conversation with the agent.

Common Questions

Deciding Whether an Agent Is the Right Tool at All

Not every use case that reaches for a SharePoint agent actually needs one. If the content in question is a handful of documents everyone already knows how to find, an agent adds a layer of AI-generated summarisation to something search already handled adequately.

Agents earn their place when a body of content is large enough that finding the right document is itself the friction, or when the value is in synthesising across several documents rather than opening any single one. A twenty item source ceiling fits that second use case comfortably. It fits the first one, a huge unstructured archive, far less well, since twenty items will rarely be enough to represent something genuinely large without careful folder-level organisation first.

The Short Version