Copilot Deep Research Tested: Real Results, Flaws, and Formatting Fails

Office-Watch.comLast updated: 21 October 2025

CoPilot,Microsoft 365,Microsoft Office,Microsoft Office for Mac,Microsoft Word,Office 365

Microsoft’s Copilot “Deep Research” promises detailed, AI-generated reports—but does it deliver? Office Watch’s hands-on tests reveal the strengths, limitations, and frustrating formatting flaws behind Microsoft’s most ambitious Copilot feature yet. From factual errors to messy Word documents, here’s what you should know before relying on Deep Research for serious work

Copilot’s Deep Research option creates a long form report on the topic you choose. It can take a while (5-10 minutes) for the AI system to dig into the details but the result is comprehensive with references to sources.

Office Watch doesn’t rely on Microsoft’s promises, we’ve tried Deep Research to see how it really works. Our tests were mostly asking for in-depth travel advice to future destinations.

In short:

Factual errors appear even in Deep Research reports.
Reports can be too long. Consider asking for a briefer (and non-quota) option first to see if that’s enough.
Don’t bother with the PDF version, any Deep Research report will need human editing.
The Word document version has insufficient formatting and errors (recoverable)

Limited Access

Microsoft 365 consumer users with some Copilot access can choose Deep Research with a monthly limit:

Premium plan: 25 uses per month of Deep Research, Researcher or Analyst.
Personal or Family plan owner: 15 uses per month in Copilot app.

Check the Facts

Every Deep Research report we tried had some factual errors. They were minor but a reminder (yet again) that AI is NOT to be fully trusted.

This Copilot option might be better suited to ‘in-house’ uses which rely on documents and other data within an organization.

One recent example of AI misuse comes from Australia. The large accounting firm, Deloitte, charged a six-figure sum for a report that included at least 20 errors including quotes wrongly attributed to a federal court judge and references to nonexistent legal and software engineering reports.

LONG REPORTS

Deep Research reports can be big with overwhelming amount of detail. Quite possibly a lot more than you need. A typical destination report was over 20 pages long.

Better to start with Copilot’s “Think Deeper” which gives a shorter result (e.g. 8 pages vs 27 pages) or let the AI decide using the “Smart (GPT-5)” option.

Those options don’t have the same monthly limits as Deep Research and give you a lot less to read, at least for a start.

Downloading the result

Deep Research reports have a download button at the bottom of the response.

Export as PDF – Do NOT choose the PDF download option because any AI-made report is sure to need some human changes. At the very least, the formatting will need updating.

Export as Docx – a Word document that can be edited and reformatted.

Generate a podcast – a clever trick but maybe not a good idea since the response might not be completely accurate.

See my thinking

At the end of the main report, just before References, is a nerdy little link “See my thinking”.

That opens a pane which shows the AI’s progress in making the Deep Research report, which sites it visited and data collected.

Word document errors

Copilot makes a .docx file which, in our tests, generates an error when opened in Word for Windows or Mac.

Word found unreadable content in “…. docx’.
Do you want to recover the contents of this document?
If you trust the source of this document, click Yes.

Click Yes and the document opens completely. Whatever the ‘unreadable content’ is, it doesn’t matter to regular users.

Some people, taught to be wary of files downloaded from the Internet, would be understandably concerned at that warning.

Does anyone at Microsoft test the output files of their own AI system? Is it naïve to expect a Microsoft product can create a file that’s compatible with another common Microsoft product?

Word document formatting

The Word document made by Copilot is just OK but lacks some of the features you’d expect and isn’t consistent. Ideally, an AI made Word document would include all the styles needed to let the human customer reformat easily and quickly.

That means customers must go through a Copilot made Word document doing basic formatting tasks. Copilot is supposed to save us from drudge work like this.

(At least the Copilot team has done a better job than the Word documents made from Microsoft’s Transcript feature, which has almost no formatting!)

The Word document uses mostly Heading 1 and Heading 2 styles. All other text uses Normal style (not Body Text). Tables use a style that can be easily changed.

Heading 3 is also used but inconsistently. Some ‘sub-heads’ use the in-built Word style but not in other parts of the same document. Those lower-level headings are just Normal style plus Bold, not Heading 3 style which is vexing, to put it mildly.

Here’s a heading “Elaborated Analysis …” that’s doesn’t use a heading style, followed by a Heading 3 line.

In other places there are a series of sub-headings under a Heading 2 but without the essential heading style.

Going through a document to add or fix heading styles isn’t hard but should not be necessary.

That means the Navigation Pane isn’t as detailed as it could be. You can add a Table of Contents for the main headings but not a more comprehensive or consistent ToC.

Horizontal lines

Copilot adds horizonal lines above Heading 2 but not as part of the Heading 2 style. That means that customers must manually find and change/remove each line instead of just changing the style.