Microsoft 365 Copilot – Persistent Source Access, Retrieval and Corpus Analysis Failures

I am posting this because I have been experiencing persistent problems with Microsoft 365 Copilot when working with academic source material, and after multiple troubleshooting attempts, support contacts, and Level 2 escalation, I still do not have a meaningful explanation or solution.

My use case is academic research.

I work with a corpus of approximately 17–20 academic textbooks, handbooks, dictionaries, and reference works in linguistics and sociolinguistics. My goal is deliberately to test the maximum practical capacity of the subscription plan I am paying for, especially regarding multi-source analysis, retrieval, source grounding, session continuity, comparison, and synthesis.

The subject matter itself is completely clear to me. I know what is contained in these sources, I am familiar with the terminology and relevant concepts, and I can independently verify whether Copilot has actually accessed the relevant material, omitted major sources, misunderstood terminology, or produced a synthesis that does not reflect the corpus.

This is therefore not a case of an unclear prompt or of a user being unable to evaluate the answer.

What I have already tested

I have tried multiple ways of providing the source material.

I have:

• used files stored within the Microsoft ecosystem;
• used OneDrive;
• tested Microsoft 365 Copilot;
• tested Copilot workflows involving Word;
• tested Notebook-style source workflows;
• rewritten and simplified prompts;
• created detailed protocols instructing Copilot to analyse all available sources;
• explicitly required source attribution;
• explicitly required comparison between authors;
• instructed Copilot not to stop at automatically generated summaries or partial retrieval output;
• converted the material into different file formats.

The file-format testing produced one particularly strange result.

Ironically, the most satisfactory results were obtained only after taking an original, fully functional academic PDF textbook — with intact indexing, tables, structured text, and no access restrictions — converting it first into an image-only PDF with no OCR, and then into an image-based PDF with OCR.

These degraded versions are clearly less suitable for serious academic analysis than the original structured PDF. Nevertheless, they were among the versions for which Copilot was at least able to confirm that it had access to the document content.

This is difficult to explain from a user perspective and suggests that the document retrieval or ingestion behaviour may not be functioning consistently.

Main problems observed

1. Incomplete corpus analysis

In one test, Copilot produced an academic-looking synthesis based on only three sources from a corpus of approximately 17–20 files.

Only after being challenged did it acknowledge that it had not systematically reviewed the remaining files.

For academic work, this is a serious problem.

If the task is to analyse a corpus, using only a small subset without clearly stating that limitation makes the result unreliable.

2. Unreliable source access

Copilot sometimes appears to know that the file exists but is unable to continue analysing it.

Depending on the session, it may indicate that:

• it cannot access additional source content;
• only previously extracted content is available;
• it cannot continue beyond the currently retrieved material;
• the source is no longer usable in the ongoing task.

This makes long-form source-based research unreliable.

3. Session continuity problems

I have also observed behaviour suggesting that source or analysis sessions can become inactive or unusable.

This is particularly problematic in multi-document work, because a corpus analysis requires continuity across multiple steps and multiple sources.

4. Summary versus source-content confusion

At several points, Copilot treated a summary or summarised retrieval result as though it were effectively the available document content.

The underlying files are not summaries. They are full academic textbooks and reference works.

An automatically generated summary should not become the endpoint of the analysis if the full document is available.

I repeatedly instructed Copilot to continue retrieving and analysing the underlying source text rather than stopping at the summary layer.

Copilot itself eventually acknowledged that this would be the correct approach, but was still unable to continue consistently.

5. Weak source traceability

For academic synthesis, I need to know which source supports each major conclusion.

This remains inconsistent.

A polished paragraph is not sufficient if I cannot reliably determine:

• which source was actually accessed;
• which authors were compared;
• which statement came from which file;
• which parts are synthesis and which parts are generic model knowledge.

Source traceability is essential for academic work.

6. Task drift

In some cases, after being repeatedly reminded of the original task, Copilot began suggesting unrelated outputs such as:

• a full course;
• a dictionary;
• exam preparation;
• a 30–50 page outline;
• other academic products.

These were not requested.

The original task was still to perform corpus-based analysis of a specific linguistic topic.

7. Word editing failures

I have also encountered cases in Word where Copilot responded that it was unable to start editing the selected text and that nothing in the document had been changed.

This adds to the impression that the issue is not limited to one specific Copilot surface.

Example of inconsistent session behaviour

I have also seen extremely simple requests behave inconsistently within the same Copilot session.

For example, the exact same question:

“What is the current time?”

first received:

“Sorry, I wasn't able to respond to that.”

and shortly afterwards received a correct current-time response.

This is obviously not an academic research task, but it is useful because it demonstrates that the instability is not always caused by complex source material.

Support history

I have already contacted multiple Microsoft support agents.

I have also interacted with Level 2 technicians.

I have submitted reports through Copilot feedback channels and through Microsoft support.

I have provided detailed explanations of the behaviour.

I have repeatedly tested the issue using different approaches.

So far, I have not received a meaningful technical explanation of the root cause, and the behaviour has not materially changed.

What I would like Microsoft to clarify

I would appreciate technical clarification on the following points:

1. What are the practical limits on the number of sources that Microsoft 365 Copilot can meaningfully analyse in one research workflow?
2. Are there documented limits on retrieval depth within large PDFs?
3. Why can a document be available in the Microsoft environment but still become inaccessible or unusable during the task?
4. What causes source-analysis sessions to become inactive?
5. Why can degraded image-based/OCR versions sometimes appear more accessible than the original structured PDF?
6. How can a user verify which files were actually accessed and analysed?
7. Why does Copilot sometimes operate only on summaries or partial retrieval output?
8. Are there differences in source access and retrieval capability between Microsoft 365 Copilot, Word, Notebooks, and other Copilot surfaces?
9. Is Microsoft aware of a wider issue involving source grounding, file retrieval, or session continuity?
10. Is there a recommended workflow for reliable academic analysis across a corpus of 15–20 large reference works?

I am not looking for another generic troubleshooting response. I would genuinely like to understand whether this is a known technical limitation, a retrieval issue, a session-management issue, or something else.

I am willing to provide examples, screenshots, test cases, and additional details if that would help reproduce the behaviour.

At the same time, I think it is reasonable to say that I cannot justify continuing to pay for a product if the core functionality I am paying for — reliable access to, analysis of, and grounding in the sources I provide — does not work consistently enough to be usable.