Which Files and Sites Can a Copilot Agent Use as Knowledge?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Which Files and Sites Can a Copilot Agent Use as Knowledge?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Which Files and Sites Can a Copilot Agent Use as Knowledge?

A Copilot agent built in Agent Builder can use up to 100 SharePoint files, one SharePoint list, 50 OneDrive files, 20 uploaded files and four public website URLs, plus Teams chats, email and connectors. To keep it to approved sources, point it at one curated SharePoint library and switch on Only use specified sources.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

That toggle prioritises your sources but doesn't block the model's general knowledge, so a hard limit needs Copilot Studio, where you can switch off ungrounded answers and web search. Two more rules shape every answer: the agent shows each person only what they're allowed to open in SharePoint, and anything you upload directly to the agent can be read by everyone who uses it.

Agent features change monthly. The numbers here are the ones Microsoft's documentation gave on 22 November 2026; if your agent's Knowledge section shows a different limit, trust the screen.

What can a Copilot agent use as knowledge, and how much of each?

A Copilot agent in Agent Builder can use public websites, SharePoint and OneDrive content, Teams chats and meetings, Outlook email, uploaded files and Copilot connectors, each with its own per-agent limit.

Small firms usually build agents with one of three Microsoft tools, and each has different limits. Agent Builder lives inside the Microsoft Copilot app and is where most small firms start. Agents in SharePoint are built from a SharePoint site by anyone with edit permission. Copilot Studio is the fuller builder for more control. The table puts their knowledge limits side by side.

Knowledge sourceAgent Builder (per agent)Agent in SharePointCopilot Studio (generative mode)
Public websites4 URLs, no more than two path levels, no query parametersNot a source25 websites
SharePoint files, folders or sites100 files20 source items in total (sites, libraries, folders, lists or files)25 SharePoint URLs
SharePoint lists1 list (up to 20,000 rows and 50 MB of text)Counts within the 20 itemsUp to 10 lists recommended per agent
OneDrive files50 filesNot listedFiles and folders supported, with separate limits
Files uploaded from your device20 filesNot listedUploaded documents supported
Teams chats and meetings5 chats and 5 meetings, or all of yoursNoNot covered here
Outlook emailYour whole mailbox, no scopingNoNot covered here
Copilot connectorsNo documented limit; an admin must enable themNoSupported

Licences decide which rows you can use at all. Copilot Chat users without a paid licence can use agents grounded in instructions or public websites at no extra cost. Grounding in SharePoint, connectors or other company data needs either a Microsoft 365 Copilot licence or pay-as-you-go billing, and for SharePoint-based agents Microsoft says the billing policy must be connected to the SharePoint agents service, with users in the security group assigned to it. What those metered answers cost is covered in how Copilot agent credits and metered billing work.

Two smaller sources are easy to overlook. People data, such as names, job titles and reporting lines, is on by default for agents created by licensed users and can be turned off with the toggle for referencing people in your organisation on the Configure tab. And OneNote can be used one page at a time through the file picker, not as a whole notebook.

Which file types can a Copilot agent read from SharePoint?

A Copilot agent in SharePoint reads DOC, DOCX, PPT, PPTX, XLSX, Loop, Fluid, PDF, TXT, RTF, ASPX, HTM, HTML, ODT and ODP files, according to Microsoft's FAQ on agents in SharePoint.

The same FAQ lists images, meeting recordings and other videos, and OneNote notebooks as coming soon, and says pages from the Site Pages library can't be added as a source for an agent in SharePoint. For files uploaded straight into Agent Builder, Microsoft gives separate size limits: 512 MB for DOC, DOCX, PDF, PPT, PPTX and TXT, and 30 MB for XLS and XLSX, with HTML supported only for SharePoint content.

A few format details decide whether answers come out right:

  • Excel works best on one sheet. Microsoft says agents respond best to Excel data held in one sheet of a workbook. A price workbook with twelve monthly tabs invites answers from the wrong month.
  • Lists have hard edges. A SharePoint list used as knowledge can hold up to 20,000 rows and 50 MB of raw text before the agent truncates it, and the Attachments column isn't indexed, so a contract attached to a list row is invisible to the agent.
  • A site doesn't include its lists. Selecting a site as knowledge leaves out the lists inside it; add the list by its own URL.
  • Some protected files fail quietly. Uploaded files with double key encryption, password protection, user-defined permissions or extract rights disabled either aren't used or stop the agent being created, sometimes without an error message.
  • Links inside documents aren't followed. In Copilot Studio, Microsoft says hyperlinks in a registered document or page aren't crawled; a linked file only counts if you add it as its own source.

If turning a multi-tab price workbook into one clean sheet is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

Does a Copilot agent show people files they can't open?

A Copilot agent doesn't show people SharePoint or OneDrive content they can't open; it respects existing permissions and sensitivity labels, so two colleagues asking the same question can get different answers.

Microsoft's SharePoint agent FAQ puts it plainly: an agent uses the sources included in it that you have permission to access, and if you lack permission to specific content, the agent doesn't show information from that content. In Copilot Studio the agent answers on behalf of the person chatting, and someone without at least Read permission gets "no response". This permission trimming is what makes agents safe to share, and it is also the most common cause of "the agent works for me but not for her" complaints.

There are three exceptions to keep in mind:

  1. Uploaded files are shared with every agent user. Files you upload from your device become embedded content in the agent. Microsoft's caution is direct: any user with access to the agent also has access to the information in the file. Information Barriers don't apply to embedded files either.
  2. Email stays personal. If you add My emails as knowledge, the agent uses your whole mailbox, but colleagues you share the agent with don't get your email as knowledge.
  3. Some tenant settings switch SharePoint off as a source. If Restricted SharePoint Search is enabled, SharePoint can't be used as knowledge in Agent Builder, and Copilot Studio blocks it too.

Permissions therefore have to be right before the agent goes live. If your sites have years of "everyone" links, start with cleaning up SharePoint permissions before turning on Copilot; an agent will happily answer from an overshared file for anyone who can technically open it.

Setting up an approved-sources library for a small firm

The cleanest way to keep an agent to approved material is to give it exactly one place to look. Instead of pointing the agent at five busy team sites, create one SharePoint site that holds only current, checked documents, and make three people responsible for what goes in it.

  1. Create the site. Make a private SharePoint team site called something like "Agent Sources". Give the whole firm Read access and only the content owners Edit access.
  2. Create one library for approved files and another for the archive. Keep superseded files in a separate "Archive" library, not in a subfolder of the approved one: the agent searches a URL and everything beneath it, so an archive folder inside the approved library is part of the agent's knowledge.
  3. Add metadata columns. Give the approved library two columns, Owner and Review by, so stale files show up when you sort.
  4. Name files so answers cite them clearly. A file called "PRICE - Backup licences - 2026-11.xlsx" tells a reader the month at a glance when the agent cites it.
  5. Check the count. Keep the approved library within Agent Builder's 100-file limit, or plan for Copilot Studio if it will grow past that.
  6. Build the agent. In the Microsoft Copilot app, choose New agent in the left pane, then Skip to configure.
  7. Add the library as knowledge. In the Knowledge section, use Enter URL and paste the library's address including the /sites/ path and without query parameters, or use Browse to pick it with the Attach cloud files picker.
  8. Restrict the answers. Turn on Only use specified sources and leave Search all websites off. Add up to four public URLs only if staff genuinely need a vendor's own page.
  9. Write the instructions using the template below, then check the Configure tab lists every source and none shows Preparing.
  10. Test on the Try it tab with the question sheet further down before you share the agent.

The instructions do as much work as the toggles. This template keeps answers tied to the library and tells the agent what to do when the files are silent:

You answer questions from [firm name] staff using only the files in
the Agent Sources site.

Rules:
- Base every answer on the approved files and name the file you used.
- Quote prices, dates and limits exactly as written in the file,
  including the month in the file name.
- If two files disagree, say so and name both. Do not pick one.
- If the files don't answer the question, reply:
  "I can't find that in the approved files. Please ask
  [operations manager]."
- Don't answer from general knowledge about our vendors, products
  or prices, even if you think you know.
- Keep answers under 120 words unless asked for detail.

A financial adviser's practice would fill the same template differently: "the approved files" become the current compliance manual, fee schedule and product panel list, and the fallback names the compliance officer rather than the operations manager. The structure stays the same.

If you want help fitting an approved-sources library into how your team already stores files, I offer one-to-one AI consultations for small businesses.

How do you keep a Copilot agent to approved sources only?

A Copilot agent stays closest to approved sources with Only use specified sources on, web search off and one curated library as knowledge; Copilot Studio can also block ungrounded answers outright.

In Agent Builder, Microsoft describes Only use specified sources as a priority, not a wall. With it on, the agent answers simple questions that need no lookup from general knowledge, uses your sources for anything that needs searching, and replies with a fallback message when your sources have nothing relevant. Microsoft states that Agent Builder doesn't support blocking general AI knowledge, and points to Copilot Studio for stricter control.

ControlAgent BuilderCopilot Studio
Prefer your sourcesOnly use specified sources toggle on the Configure tabKnowledge sources added at agent level
Block answers without a sourceNot supportedTurn off Allow ungrounded responses in Generative AI settings
Keep the open web outLeave Search all websites offTurn off Web Search / Use information from the web
Narrow by date or authorPoint at a narrower folderFilters on Title, Author, Modified by or Modified on
Mark a source as trustedNot availableOfficial source (classic orchestration only)

Copilot Studio's strict setting has a quirk worth knowing before you choose it. With Allow ungrounded responses off, the agent returns an answer only when it includes an in-text citation to a source. Microsoft notes that the model sometimes produces a correct answer without a citation, and the agent then withholds it as if nothing was found, so the same question can work on a second try. Microsoft's fix is an instruction such as "Always include an in-text citation to the source document for every statement", and avoiding instructions that strip references or force rigid formats.

The Modified on filter is useful for small firms with messy libraries. A filter of "Modified on, on or after" a date six months ago keeps the agent off files nobody has touched since spring, while you move them to the archive properly.

How quickly does a Copilot agent see new or changed files?

A Copilot agent picks up newly added SharePoint and OneDrive files within up to several minutes, showing Preparing beside them until they're ready, and it follows renames and deletions of files it already uses.

You can test the agent while a source is still preparing, but answers won't use that file until the status clears; the refresh button at the top of the Knowledge section reloads the status. Uploaded files behave differently because they are embedded copies: when the original changes, remove the old upload with the X next to it and upload the new version, or better, keep that file in SharePoint and point the agent at it there.

Structural changes need more care than file edits. In Copilot Studio, renaming the SharePoint site or folder behind a knowledge source can break the link and leave the agent unable to reach its knowledge until you add the new link. Agents in SharePoint handle hub sites generously: since September 2025, selecting a hub site as a source includes its associated sites, but an agent created before then needs the hub site removed and added again. A newly created site can take several minutes to appear in the file picker, and communication sites may show only under Quick Access and Recent.

A sensible routine for a small firm: content owners replace files in the approved library as soon as a price or policy changes, move the old version to the archive the same day, and ask the agent one test question about the change before telling staff it's live.

Wrong-answer tests to run before staff rely on the agent

Ten minutes of deliberate attempts to break the agent saves weeks of quiet wrong answers. Write a short question sheet with the answer you expect and the file it should come from, then run it on the Try it tab, once as yourself and once as a colleague with narrower permissions. Here is a filled-in sheet for an illustrative software reseller:

#QuestionExpected answer and sourceWhat it tests
1What does 25 seats of the backup product cost a year?Figure from "PRICE - Backup licences - 2026-11"Uses the current price file, not an older one
2How many days do customers have to cancel a new subscription?Period from the terms-of-sale documentQuotes the policy exactly
3What discount do we give charities?"I can't find that in the approved files" (no such policy)Fallback instead of an invented discount
4What is a competitor's list price for the same product?Fallback; no web searchStays off the open web
5What is our sick-pay policy? (asked by a sales rep)No answer from the HR folder the rep can't openPermission trimming
6Which file covers renewals for the backup product?Names the renewals procedureCites a file, as instructed
7What did the old price list say for 25 seats?Fallback; the archive is out of scopeArchive kept out of knowledge
8Summarise the onboarding steps for a new reseller customerSteps from the onboarding procedure, in orderLong-document handling
9Does the security product support single sign-on?Answer from the vendor's product guide PDFReads PDFs correctly
10Two files give different support hours. Which is right?Names both files and the conflictFollows the "if two files disagree" rule

Question 1 is where most first runs fail. Here is an illustrative exchange and the fix:

Question: What does 25 seats of the backup product cost a year?

Agent (illustrative output):
"25 seats of the backup product cost $1,140 a year
(PRICE - Backup licences - 2026-05.xlsx)."

The figure came from May's price list, which someone had saved into a subfolder of the approved library instead of the archive. The fix took two minutes: move the May file to the Archive library, wait for the agent to pick up the change, and ask again. The answer then cited the November file. Running question 7 alongside it confirms the old file is really out of reach.

Question 5 should "fail" in the right way. If the sales rep gets nothing back about sick pay, permission trimming is working; if the rep gets an answer, the HR folder is overshared and the problem is SharePoint, not the agent.

Worked example: a 20-person software reseller's knowledge agent

Here is the whole set-up for one illustrative firm. A 20-person B2B software reseller wants an internal agent that answers staff questions about prices, terms and product features. Eight people hold Copilot Business licences; the other twelve use Copilot Chat. Product information currently sits across four team sites, two personal OneDrives and a shared mailbox.

The library. The operations manager creates the Agent Sources site and moves in 64 files: 12 policy documents, 6 price workbooks rebuilt as one sheet each, 30 vendor product guides as PDFs and 16 procedures. That is well inside the 100-file limit. Older price lists, 41 files in all, go to a separate Archive library. Two vendor licensing pages are added as public URLs, both within two path levels.

The time. As a rough planning figure for a firm this size, gathering and renaming the files takes about three hours, rebuilding the price workbooks another two, building the agent half an hour, and the first test run an hour.

The first test run. Seven of the ten questions pass. Question 1 fails on the May price file left in a subfolder. Question 3 fails because Only use specified sources was left off, and the agent invents a 10% charity discount that the firm has never offered. Question 10 fails because the agent picks one set of support hours without mentioning the other file. After moving the file, turning the toggle on and adding the "if two files disagree" line to the instructions, the second run passes all ten.

The access question. For the twelve Copilot Chat users, a SharePoint-grounded agent needs pay-as-you-go billing connected to the SharePoint agents service, with those users in the billing policy's security group. The owner sets a monthly budget with an alert at 50%, gives the agent to the eight licensed users first, and decides on the other twelve after a month of real usage.

The upkeep. Each content owner gets a monthly reminder to sort the library by the Review by column, archive anything superseded, and ask the agent one question about each changed file. If you plan to give the agent to more teams later, the setting that controls who can use agents at all is explained in the admin guide to agents disabled in your organisation.

What goes wrong with agent knowledge in a small firm?

Agent knowledge usually goes wrong through old files left inside the agent's scope, confidential files uploaded instead of linked, and sources that silently fail because of a URL, file type or label.

  • The archive inside the approved folder. It shows up as answers citing last spring's prices. Keep archives in a separate library; the agent searches everything under the URL you give it.
  • The salary file uploaded "just for testing". It shows up months later when a junior asks the agent about pay. Uploaded files are readable by every agent user, so test with SharePoint links instead and delete test uploads.
  • The URL with a query string. A link copied from a filtered library view contains "?" parameters and is rejected or points at the wrong thing. Copy the plain library address, or use Browse.
  • The 101st file. The library grows past Agent Builder's 100-file limit and newer files never become knowledge. Watch the count, split agents by topic, or move to Copilot Studio.
  • The list attachment. A supplier contract attached to a SharePoint list row never appears in answers, because the Attachments column isn't indexed. Store the contract in a library and add that.
  • The general-knowledge answer. With Only use specified sources off, the agent fills gaps with plausible guesses about your products. Turn it on, add the fallback line to the instructions, and keep question 3 on your test sheet.
  • The agent nobody reviews. Six months on, the agent still cites files whose owners left. An admin can hand ownerless Agent Builder agents to the former owner's manager with an agent management rule, but the files need owners too; the Owner column in the library is the cheapest safeguard.

If your firm is still deciding what belongs in a shared knowledge store before any agent reads it, building a company knowledge base AI can answer from covers that groundwork, and building a leave-policy agent without a licence shows a single-document agent from start to finish. Microsoft's full list of sources and limits is on its Agent Builder knowledge page on Microsoft Learn.

Copilot agent knowledge: follow-up questions

Can a Copilot agent use a whole SharePoint site as knowledge?

Yes. In Agent Builder, enter the site's URL including its /sites/ path and without query parameters, and the agent searches that URL and its subpaths. Microsoft lists the SharePoint limit as 100 files per agent. A site doesn't bring its lists with it, so add a list separately by its own URL, and remember the agent still shows each person only what they can open.

Can I stop a Copilot agent from using general AI knowledge?

Not completely in Agent Builder. Its Only use specified sources toggle makes the agent prioritise your sources and give a fallback message when they lack an answer, but Microsoft says it doesn't block general knowledge. For a harder limit, build in Copilot Studio and turn off Allow ungrounded responses and web search, which blocks answers given without a knowledge source or tool.

Do staff need a Copilot licence to use an agent grounded in SharePoint?

They need either a Microsoft 365 Copilot licence or pay-as-you-go billing. Copilot Chat users can use agents grounded in instructions or public websites at no extra cost. For agents that use SharePoint files, Microsoft says the pay-as-you-go billing policy must be connected to the SharePoint agents service and users must be in the security group assigned to that policy.

Are files uploaded to an agent visible to everyone who uses it?

Yes, in effect. Microsoft warns that anyone with access to the agent also has access to the information in files uploaded to it, because uploaded files become embedded content rather than staying under their SharePoint permissions. Keep confidential material in SharePoint, where the agent respects each person's access, and upload only files every agent user may read.

Why does a file show Preparing in the agent's Knowledge section?

Preparing means the agent hasn't finished processing a newly added SharePoint or OneDrive file. Microsoft says new uploads can take up to several minutes to be included, and you can test the agent meanwhile, but answers won't use that file until it's ready. Use the refresh button above the Knowledge section to reload the status.

Further reads

Sources: Microsoft Learn pages on adding knowledge sources in Agent Builder, agent capabilities and licensing models, the Copilot Studio knowledge sources summary and adding SharePoint as a knowledge source; Microsoft Support's FAQ on agents in SharePoint. All read 22 November 2026.

Want a second pair of eyes on your setup?

Bring the process you're trying to fix. On a 1:1 call we'll work out which AI tools should handle it, which parts stay manual, and how to roll it out without disrupting the team.

Book a 1:1 call with me