Teams that already pay for Microsoft 365 Copilot and Copilot Studio often hit the same wall. The agent gives good answers when someone uploads a few files into the chat, then gives thin or wrong answers once it is pointed at the SharePoint folders where those same files live.
The model has not got worse. Uploading and linking are two different ways of finding information, and SharePoint content needs preparing before an agent can find the right file. This guide explains the difference, what you can build on the licenses you already have, and when you need more than Copilot alone.
What is actually different between uploading a file and linking a SharePoint folder?
Building an agent on SharePoint documents?
Agnotic Technologies builds SharePoint document AI agents and the retrieval behind them, so each answer comes from the right file and cites it.
An uploaded file is copied into the agent's own knowledge store, so the agent reads that exact document. A linked SharePoint folder is searched when the question is asked, so the agent only sees the passages the search returns, from whichever files it ranks highest.
Microsoft's Copilot Studio documentation describes the two paths separately: uploaded documents are stored in Dataverse, while a SharePoint knowledge source connects to a SharePoint URL and uses Graph search to return results.
- Ranking decides what the agent reads: in a folder of similar files, such as one report per vessel or per project, the search can return a passage from the wrong one.
- File size matters: Microsoft states that without a Microsoft 365 Copilot license in the same tenant as the agent, generative answers can only use SharePoint files under 7 MB, and that files up to 200 MB need that license with tenant graph grounding turned on.
- Permissions apply per person: the agent only shows SharePoint content the person asking can open, so two people can get different answers to the same question.
- Folders drift: files are added, renamed and copied over time, while an uploaded set stays fixed.
When an agent works on uploads and fails on links, the cause is almost always retrieval, not the language model.
We already pay for Microsoft 365 Copilot. Can we build the agent on what we have?
Often, yes. If the questions are about a well-organised set of Word, PowerPoint and PDF files that every user can already open, a Copilot Studio agent with SharePoint as its knowledge source is a sensible first build.
Check the license position first, because, as above, it decides how large a SharePoint file the agent can use. Then check the fit:
- Mostly text: the documents are Word, PowerPoint or PDF files with real text, not scans or drawings.
- Short answers: each question can be answered from one or two passages.
- Access in place: users can already open everything the agent should use.
- A person checks: answers are read by someone, not passed straight into another system.
Copilot Studio agent development is quick to prototype for cases like these. Add custom retrieval only when testing shows the search returning the wrong files.
How should we organise SharePoint folders so an agent finds the right vessel or project?
Give every file one clear home and a name the search can match. An agent finds the right vessel or project when the folder path, the file name and the metadata all say which vessel or project a document belongs to, and when old copies are out of the way.
Microsoft notes that the agent searches only the folder or site URL you register and its subfolders, so the structure you register is the boundary of what the agent can find.
- One folder per vessel or project: put the identifier in the folder name and in each file name, so a search for one cannot match another.
- Metadata columns: add columns such as vessel or project, document type and date, and fill them when files are uploaded.
- Archive old versions: move drafts and superseded revisions to a library the agent is not pointed at.
- Register narrow sources: point the agent at the folders it needs, not the whole site.
- Test with real questions: run the questions users ask and check which file each answer came from.
This tidy-up is usually the cheapest way to improve Copilot agent SharePoint accuracy, and it helps people find files too.
When do we need a vector store or Azure AI Foundry instead of Copilot alone?
You need your own retrieval layer when search over SharePoint cannot reliably find the right passage. That usually means many near-identical documents, scans and tables that need extraction first, or answers that must combine facts from several files.
- Many similar documents: reports that share one template, where the agent must filter by vessel, project or date before it searches.
- Content search cannot read well: scanned pages, photos of forms, drawings and complex tables.
- Answers that need structure: comparing findings across reports, or producing a summary in a fixed format.
- Control over citations: every answer links back to the file and page it came from.
In that design, documents are extracted, split into passages, tagged with metadata and indexed in a vector store such as Azure AI Search, and the agent, built in Azure AI Foundry or Copilot Studio, queries that index. RAG on SharePoint documents keeps SharePoint as the place people work, while the index becomes the place the agent searches.
The same pattern supports vessel inspection report automation, where an agent reads inspection reports and drafts a summary for a person to check.
Need retrieval you can control?
Agnotic Technologies designs the extraction, index and retrieval layer behind document agents, with metadata filters and citations back to the source file.
What should the agent do with files it cannot read?
The agent should say so and list them. An agent that silently skips a file gives an answer that looks complete but is not, which is worse than no answer at all.
Some files will not be readable: scanned pages with no text layer, password-protected or labelled files, very large files and formats the knowledge source does not support. Plan for them from the start.
- Keep a skipped-file log: record each file that could not be read and why.
- Tell the user: when an answer depends on a folder with skipped files, the agent names the files it could not use.
- Route fixes to an owner: someone converts, removes the password from or replaces the file, and it is indexed again.
- Never guess: if the file that holds the answer was not read, the agent says it cannot answer instead of filling the gap.
Where should a team start?
Start with one document library and a list of real questions your users ask. Run those questions against uploaded copies and against the SharePoint link, compare which file each answer came from, and fix the gap you can see.
- Pick one library: one vessel class, one project type or one department.
- Collect real questions: note the file and page that holds each correct answer.
- Tidy that library: one home per file, clear names, metadata and old versions archived.
- Test in Copilot Studio: use SharePoint as the knowledge source and record where each answer came from.
- Decide on retrieval: if the right file still is not found, add a custom index for that library only.
Agnotic Technologies builds these agents for clients, from a Copilot Studio agent on a tidy library to a custom retrieval layer with citations. Each project starts from one library and the questions people actually ask.
Start with one library
Share one SharePoint library and the questions your team asks, and we will show where the agent loses the right file and how to fix it.
Related reading on Agnotic
More on scoping a first AI agent, retrieval, document automation and choosing a build partner.
