Connecting AI Engines to Company Knowledge

Microsoft Copilot and SharePoint: Why It Can't Find Your Documents, and How to Fix It

By the Praxiron team · Last updated July 5, 2026 · 12 min read

Why does Copilot miss documents that are right there?

Because between your question and your document library sit four checkpoints, and a document that fails any one of them never reaches the answer. Per Microsoft’s documentation, the checkpoints are: the search index, since recently uploaded documents may not be indexed yet and only modern SharePoint pages are supported; permissions, since Copilot runs on delegated permissions and a user without read access gets nothing back; retrieval limits, since generative answers over SharePoint use only the top three search results; and file limits, since size caps and format support decide what Copilot can process at all. There is also a fifth, quieter case: content held for moderation returns no response, with no indication that anything was filtered.

The frustration is legitimate because the document usually is right there. You can open it in the browser. Classic search finds it. A colleague quotes it in a meeting. But Copilot is not browsing your library the way you do; it composes an answer from a small set of retrieved results, and each cause below describes a way a document silently drops out of that set. What follows is a cause-by-cause walkthrough with the fix for each, and then the part that matters more: what no amount of retrieval tuning fixes when the answers feed real decisions.

Cause 1: indexing lag and unsupported page types

Copilot can only draw on what the search index contains, and the index is not instantaneous. Per Microsoft’s documentation, recently uploaded documents may not be indexed yet, which means a file added this morning can be fully present in the library and fully absent from Copilot’s view of it. Teams hit this constantly during exactly the moments they most want AI help: a new policy lands, a contract is uploaded before a call, and Copilot answers as if the file does not exist, because from the index’s point of view it does not exist yet.

Page types are the second half of this cause. Microsoft’s documentation states that only modern SharePoint pages are supported. Organizations that migrated over the years often carry a long tail of classic pages, wiki pages, and legacy layouts. The content on those pages is readable by any human who opens them and invisible to the generative-answer pipeline. If your intranet’s most authoritative guidance lives on a classic page that nobody has rebuilt since the migration, Copilot will keep answering around it.

The symptom pattern for this cause: Copilot misses new content and old content, while handling the middle of your library fine.

Cause 2: permissions (read access required, and the oversharing flip side)

Copilot operates under delegated permissions. In plain terms, per Microsoft’s documentation, it acts as the user who is asking: if that user does not have read access to a document, nothing from that document comes back, and Copilot behaves as if the document does not exist. This is the correct security posture, and it is also a common source of “why can’t it find the file” tickets. The person who built the agent or configured the knowledge source can see the file, so it works in testing. The person asking in production cannot, so it fails for them, silently, with no message saying access was the reason.

The flip side deserves equal attention. Because Copilot faithfully inherits whatever permissions exist, it also inherits every permission mistake. Most tenants carry years of accumulated oversharing: site-wide links created for convenience, groups nobody has pruned, folders shared with “everyone” during a crunch and never revisited. Before Copilot, that oversharing was buried under the effort of manually finding things. After Copilot, everything a user can technically reach becomes instantly discoverable by asking a question in plain language. Microsoft’s own deployment guidance recommends reviewing sharing settings and access before broad rollout for exactly this reason. The permission surface did not change; the ease of traversing it changed completely. We cover this dynamic in depth in what actually happens to your permissions when you connect a file server to AI.

The symptom pattern: results differ by person. One user gets a solid answer, another gets nothing, and the difference maps to access, not to the question.

Cause 3: retrieval limits (generative answers use only top search results)

This is the cause most people have never heard of, and it explains the largest share of “the document is indexed, I can access it, and Copilot still ignored it” cases. Per Microsoft’s documentation, generative answers over SharePoint use only the top three search results to compose the response. Not the top thirty. Three. Copilot does not read your library; it reads a very short shortlist produced by search ranking, and everything below the cut line contributes nothing.

Two consequences follow directly. First, ranking becomes destiny. If the decisive document ranks fourth for the way you happened to phrase the question, the answer is built without it, even though the document exists, is indexed, and is readable by you. Ask again with different wording and a different shortlist forms, which is one reason the same question can produce different answers on different days. Second, per Microsoft’s documentation, when the search returns no results there is no response at all. Silence. And silence is ambiguous in the worst way: it looks identical whether the knowledge genuinely does not exist in your company or the retrieval simply failed to surface it.

The symptom pattern: answers that use some relevant documents but not the best one, answers that change with rephrasing, and questions that return nothing even though you know the material exists.

Cause 4: file size and format limits

Large files are a documented boundary. Per Microsoft’s documentation, file size limits apply to what Copilot can process: 200 MB for SharePoint files with a Microsoft 365 Copilot license and Enhanced search results enabled, and 512 MB for uploaded files. Those ceilings sound generous until you meet the real population of enterprise documents: scanned drawing sets, recorded-meeting transcripts bundled with media, decade-spanning master spreadsheets, appendix-heavy PDF reports. A file over the limit does not produce an error a user would recognize as a size problem; it simply fails to inform answers.

Format matters as much as size. Content that is technically inside a supported file can still be practically unreachable: text baked into scanned images without OCR, information locked in complex table layouts, or data living in embedded objects. The file passes the size check and the substance never makes it into the answer.

Finally, the moderation case belongs in this family of silent misses: per Microsoft’s documentation, content held for moderation returns no response, without indication. From the user’s chair it is indistinguishable from every other silence in this list.

The symptom pattern: consistent misses on the same big or unusual files, regardless of who asks or how the question is phrased.

How to fix each cause, step by step

Work the causes in order. They are ordered roughly by how often they are the culprit and how cheap they are to check.

Fix for indexing lag and page types

1. Check recency first. If the missing document was uploaded or heavily edited in the last day, wait and retest before touching anything else. Per Microsoft’s documentation, recently uploaded documents may not be indexed yet; a large share of “Copilot is broken” reports are just the index catching up.
2. Confirm the site and library are searchable. In SharePoint site settings, verify that the site and the specific library are not excluded from search results. An exclusion set years ago for a tidy intranet will now also hide content from Copilot.
3. Test with classic search. If SharePoint search itself cannot find the document, Copilot never will. Fix search visibility first; Copilot inherits the result.
4. Rebuild critical classic pages as modern pages. Only modern pages are supported, so inventory the classic and wiki pages that hold decision-relevant guidance and migrate that content. Prioritize by how often the content is asked about, not by page age.

Fix for permissions

1. Reproduce as the affected user, not as an admin. Most permission misses are invisible to the person doing the troubleshooting because the troubleshooter has broader access.
2. Verify read access at the document level, then the library, then the site. Broken inheritance in the middle of that chain is a classic cause.
3. For Copilot Studio agents using SharePoint as a knowledge source, remember the delegated model: the agent answers with the asking user’s permissions, so “it worked when I tested it” proves nothing about other users.
4. Then run the oversharing pass in the other direction: review sharing links, “everyone” grants, and stale group memberships before expanding rollout, as Microsoft’s deployment guidance recommends. Fixing under-access without fixing over-access solves the ticket and leaves the exposure.

Fix for retrieval limits

1. Improve the ranking of decision-critical documents: accurate titles, descriptive first paragraphs, and consistent metadata. You cannot raise the top-three ceiling, so the practical lever is making sure the right documents occupy those three slots.
2. Consolidate duplicates and archive superseded versions. Every near-duplicate competes with the authoritative copy for the same shortlist, and ranking does not know which one your organization considers current.
3. Ask narrower questions scoped to specific sites or topics where possible; a smaller candidate pool gives the right document better odds of ranking.
4. Accept what tuning cannot do: three results is the documented ceiling for generative answers over SharePoint. Curation improves the odds; it cannot make retrieval read the fourth document.

Fix for file size and formats

1. Check the missing file against the documented limits: 200 MB for SharePoint files with a Microsoft 365 Copilot license and Enhanced search results, 512 MB for uploads.
2. Split oversized files where the content allows. A 400 MB “everything” PDF becomes findable as chapters.
3. Convert content-bearing scans with OCR so the text actually exists as text.
4. Extract decision-relevant tables and appendices into their own clean documents. Retrieval favors focused files over sprawling ones anyway, so this fix also helps with Cause 3.

Run through all four fixes and Copilot will find dramatically more of what is really there. That is worth doing on its own terms, and Microsoft has built a genuinely capable retrieval pipeline once it is fed properly. Which brings us to the harder question.

You fixed retrieval; the reasoning gap remains

Everything above makes Copilot better at fetching documents. None of it makes the output something you can safely act on for decisions that carry real cost, because retrieval and reasoning are different jobs, and the second one was never in the pipeline.

Consider what a perfectly tuned Copilot still does not do. It does not apply your company’s decision rules: which document wins when two conflict, which source is authoritative for which question, what recency rules govern which policy. It composes fluent text from the shortlist and does not clearly separate what your documents say from what the model concluded, so a reader cannot tell where evidence ends and synthesis begins. It attaches the same confident tone to a well-supported answer and a thinly supported one. And its failure mode is silence: per Microsoft’s documentation, no results means no response, which leaves the person asking with no way to distinguish “our company has no answer to this” from “retrieval missed it.” Those are opposite situations demanding opposite actions, and they look identical on screen. The same structural gap produces the day-to-day trust problem of the same question getting different answers, since a ranking-dependent shortlist plus unseparated synthesis varies in ways no user can predict.

“Silence is the worst failure mode for a decision-maker. When a tool returns nothing, you cannot tell whether the knowledge does not exist or the retrieval missed it. An explicit abstention, with the reason attached, is information you can act on. Silence is not.”

The Praxiron team

This gap, not model quality, is where enterprise AI initiatives stall. MIT NANDA 2025 found that 95% of enterprise generative AI pilots showed no measurable P&L impact. PwC’s 2026 Global CEO Survey put numbers on the executive view: 56% of 4,454 CEOs report no cost or revenue improvement from AI in the past 12 months. And S&P Global Market Intelligence 2025 found 42% of companies abandoned most of their AI initiatives. Retrieval that works and reasoning that is absent is precisely the pattern behind those numbers: the tool demos well, the outputs cannot be verified cheaply, senior people re-check everything, and the gains evaporate. The full argument for why fetching text is not the same as supporting a decision is in RAG isn’t enough.

None of this is a flaw in Copilot’s mission. Microsoft built Copilot to retrieve and synthesize across Microsoft 365, and after the fixes above it does that well. The opportunity is what becomes possible once that foundation is in place: a tuned Copilot plus a layer that adds the reasoning, calibration, and honesty that retrieval alone was never designed to provide.

What a knowledge and control layer adds above Copilot

A knowledge and control layer sits between your company’s knowledge and the AI engines, and it changes what an output is, not just where it comes from.

It starts with decision DNA: your standards, precedents, and the judgment of your senior experts structured deliberately, with the hierarchy and recency rules that decide which source wins, instead of a flat library where ranking decides. On top of that structure, every output carries a source reference showing exactly which documents it rests on, with document content separated from generated conclusions, so checking an answer takes minutes instead of redoing the work. Each output also carries calibrated confidence, a level that visibly drops when support is thin rather than a uniformly assured tone. And when the sources are insufficient, the layer abstains explicitly: “no sufficient source” replaces the ambiguous silence you met in Cause 3, telling the decision-maker exactly where the company’s knowledge ends.

Control extends to access. Instead of inheriting storage permissions as they stand, oversharing included, the layer governs access by file type and role: financial documents reachable by finance roles, HR files by HR, drawings by engineering, independent of how the underlying SharePoint sharing accumulated over the years.

The layer is also engine-agnostic by design. It works above Copilot, and equally above ChatGPT, Gemini, Claude, and Grok, so the knowledge you structure is an asset your company owns rather than a configuration of one vendor’s tool. If your team is also connecting ChatGPT to company files, the same governed knowledge serves both engines instead of being rebuilt per tool. Praxiron is a platform built as exactly this category: decision DNA, source references on every output, calibrated confidence, abstention when sources are insufficient, and permission control by file type and role, above every engine. If you want to see how outputs with sources, confidence, and abstention work in practice, start with how the platform works.

Copilot alone vs. a knowledge and control layer