What Copilot can do with a scanned document (and what it can't)

Working with Copilot and scanned documents in PDF It has become one of the most common questions for anyone who is Setting up agents with Copilot Studio or using Copilot within Microsoft 365. Many users assume that if Copilot "reads PDFs", it will also be able to understand a scanned file as an image... and that's where the misunderstandings begin.

In this article we will break down, calmly and directly, What Copilot can and cannot do with a scanned document, how it behaves in Word and other Microsoft products, what role OCR (optical character recognition) plays, and how to set up a realistic workflow to leverage your scanned PDFs as knowledge sources.

Can Copilot directly read a scanned PDF?

The first thing to clarify is that a A scanned PDF is not a text PDFIn a "normal" PDF, the content consists of selectable text; in a scan, there are only images of the pages, without any accessible text layer. From the perspective of any language model, if no OCR is performed beforehand, there's nothing to read.

The agents of Copilot Studio, according to current Microsoft documentation and community experience, They do not natively process text from purely image-based PDFsFor Copilot to reason about content, there must be previously extracted text. This applies both to agents using knowledge sources and to document upload scenarios in other Microsoft applications.

Some users claim their agent has "read" scanned PDFs, but more often than not, it's actually the file already had a hidden text layer (for example, because the multifunction printer or scanner automatically applies OCR) or it was a mixed PDF, with text and some scanned images. Without that text layer, Copilot has no structured content to work with.

In practical terms: if you upload a PDF that only contains scanned images and you haven't run it through OCR, Copilot Studio is not going to extract the text from nothing.You will need a preliminary character recognition step with an OCR tool, ideally integrated into your workflow in Azure or your document management system.

How to integrate documents in Copilot Studio

Copilot Studio allows you to create agents that use prompts and knowledge sources to respond to users. One of the most important new features is the ability to define instructions that establish a logic for extracting and enriching data from documents and images.

In practice, you can configure a message in Copilot Studio, Power Apps, or Power Automate specifying what you want the agent to do with the files it receives: analyze, summarize, extract specific fields, transform format or perform any other type of processing consistent with your use case.

Within these requests you can add one or more file input typesMicrosoft states that supported file types include PNG, JPG, JPEG, and PDF. This means the agent can work with documents and images, but it doesn't imply that it will automatically convert an image into structured text without the aid of a recognition process.

The real value lies in the fact that, once this processing logic is defined for Copilot agents, manual labor is greatly reduced for end users. Instead of each person having to open, copy, paste, and format, the agent always executes the same logic, ensuring consistency and saving time.

Furthermore, these prompts can reuse across different agents, applications, and automated workflowsThis means that a single investment in logic design has a multiplier effect across your organization. Currently, Microsoft indicates that these requests can be incorporated into agents through themes; actions are not yet supported in this context, although the roadmap is expanding capabilities.

Using OCR and Azure to process scanned PDFs

When your knowledge sources or working documents are scanned PDFs, the reasonable solution is to introduce a OCR step before reaching CopilotThis is where Azure AI Document Intelligence (formerly known as Form Recognizer) and similar services come into play.

With Document Intelligence you can send your scanned PDFs and extract the text and structureParagraphs, tables, headings, key fields, etc. The service is able to detect the content even if the source file only contains images and return it in a format usable by your applications.

Once you have extracted the text, you have several options: you can Generate a new PDF with a text layerYou can save the content as Word, JSON, or HTML, or even store the structured data directly in a repository like Dataverse or SharePoint. From that point on, Copilot will be able to use these documents as a reference.

A very practical strategy is to automate this flow: when someone uploads a scanned PDF to a specific SharePoint library, a Power Automate flow is triggered that calls Document Intelligence, save the processed result and incorporates it into the knowledge base accessible to your Copilot Studio agent.

With this approach, your users only see that "they upload a file and Copilot understands it," but underneath there is a whole OCR process. normalization and enrichmentThis intermediate layer is what makes the difference between an agent that fails with scanned PDFs and one that can actually answer detailed questions about them.

Applications with documents and images in Copilot Studio

Another important aspect is how the applications that accept files In Copilot Studio, Microsoft describes a feature where the agent creator defines a message and adds one or more input types: image or document, including PNG, JPG, JPEG, and PDF.

When designing these requests, you can specify the logic that Copilot will apply: for example, that it extracts certain data from an invoice, that Summarize the content of a long report, which converts the text into a table or applies a standardized format that will then be reused in other processes.

The great advantage is that, once the request is configured, All users of the agent will enjoy the same automationIt eliminates the variability of each person copying and pasting data in their own way, and avoids the tedium of repetitive operations that Copilot can do in one or two chat interactions.

From an administrative point of view, this feature is activated in a default without needing to configure anything additionalSimply create the message, choose the supported file types, and then include it in one of the Copilot agent topics.

When discussing files that may include images (such as scanned PDFs), it's essential to keep in mind that the logic you design must take this into account. What part of the job does Copilot do, and what part does OCR do?If the text does not yet exist, Copilot needs another technological layer to provide it in a structured form.

Copilot in Word: compose, transform, and chat about the document

Beyond Copilot Studio, another essential part of the puzzle is how it behaves Copilot in Word With long documents, reports, proposals, or any file you want to review or transform, Copilot in Word uses the document's content to generate drafts, summaries, and answers to direct questions.

When you open a new document or insert a blank line into an existing one, Word displays the experience of “Generate a draft with Copilot”There you can write a simple request (“write a business letter introducing this product”) or something more elaborate with outlines, notes, or even references to other files in your Microsoft 365 environment.

If you have a Microsoft 365 Copilot license, you can use the button “To reference a file” To attach up to three files as the basis for the draft that Copilot will generate, you can also type “/” followed by the file name to quickly select it. Copilot will only use the documents you choose, without accessing other organizational data that you haven't explicitly referenced.

Once the text proposal has been generated, you can preserve it, discard it, or request that it be regeneratedIt is possible to navigate between different versions, adjust the tone or length with new instructions (“make it more concise”, “use a more formal tone”), and even go back and modify the original message to better guide the result.

In documents that already contain content, Word incorporates the button “Inspire me”which allows Copilot to continue writing from what you have already written, maintaining consistency with the style and theme of the existing text.

Transform and structure content in Word with Copilot

Copilot in Word isn't just for writing from scratch, it also makes the work of rewrite and polish documents that already exist. Often, editing and restructuring takes more time than writing a new text, and this is where Copilot can save you a lot of time.

If you select a text fragment in the document, you'll see a Copilot icon in the margin. From that menu, you can choose the option to “Automatic rewriting” so Copilot can suggest alternative versions. From there, you can adjust the response, add new details, or request another batch of suggestions.

The same text selection allows you to activate the option to “View as table”Copilot takes the content and turns it into a table, which is very useful when you have lists or blocks of information that would be clearer if they were tabulated. You can then further refine the table with instructions such as "add a third empty column."

In both cases, the idea is for Copilot to do the heavy lifting: reformulating, structuring, and shape the information without having to start each change from scratch. You still have final control to decide what stays and what gets discarded.

It's important to remember that these transformations work on the text that Word can read. If you've inserted a scanned PDF into a Word document as an object or as a raw image, Copilot will not understand the content of that image unless accessible text has been previously generated (for example, by running the PDF through OCR and pasting the result as text into the document).

Chat with Copilot about the document

Another powerful feature of Copilot in Word is the chat panelIt is accessible from the Home tab of the ribbon. When you open it, you can ask questions about the document, request summaries, clarifications, additional ideas, or even request tables and lists that you can then insert into the text.

When you ask Copilot to summarize the document or answer specific questions, the system includes references with citations to the parts of the text from which the information was taken. This is especially useful in long or technical documents, because you know exactly which paragraphs it was based on.

The chat is not limited to a single interaction; you can ask back-and-forth conversationsRefining the response, requesting more detail, or asking for a different perspective. Additionally, chats are stored in the Copilot history, allowing you to return to previous conversations and continue working where you left off.

Again, the underlying requirement is that there is sufficient textual content so that Copilot can work. Microsoft indicates that, in order to generate summaries, the document must have at least 20 reference words, and that there are length limits, for example, around 1,5 million words or about 300 pages for detailed summaries.

If the document includes graphics, images, or embedded files, Copilot will attempt to use the available textual information, but It will not reconstruct complex information that is only in a scanned image. without additional support.

Automatic summaries and level of detail control

In many cases, especially with long reports, Word may display a automatic summary at the top of the document when you open it. This feature, available to those with a Microsoft 365 Copilot license who save the document to OneDrive or SharePoint, provides a quick overview of key topics.

If you're reviewing a long file that was just sent to you, this summary lets you get a general idea without having to read it from beginning to end. As an author, it also helps you check if your content is organized as you intended and if the main messages are being conveyed effectively.

The summary includes an option for Open it in the chat and ask personalized questions. From there, you can request more details on a specific point, have them restructure the information, or prepare a text to share with other teams.

The tab "Discussion" It allows you to see a summary of the document's conversations and comments, so you don't have to go through each annotation individually. Copilot displays the key topics discussed and their associated threads, significantly speeding up collaborative review.

You can also consult the section on “Conclusions” or key figuresThis section highlights important numbers from the document or examples of questions you could ask Copilot. This is very useful when you're looking for specific data ("earnings per share," "growth figures by region") without knowing exactly where it is in the text.

When the summary does not appear and other limitations

You won't always see an automatic summary when you open a document. If the file isn't saved to OneDrive or SharePoint, or you don't have the appropriate license, Word will display an option to... manually generate a summary instead of displaying it directly at the top.

If nothing appears, you can always open the Copilot panel from the ribbon and type a request. “summarize this document”Copilot will analyze the text and provide a summary with a references section to review the source citations.

Microsoft indicates some current limitations: for example, the need for a minimum of content To make it worthwhile to generate a summary, there is an approximate length limit for detailed summaries, and although Copilot takes the entire document into account, it does not always provide citations to very advanced content within the text.

In any case, these summaries still depend on the document containing text readable by the systemAn attached image or scanned PDF file inserted without OCR will not be able to be analyzed in depth. In these cases, the best practice is to first run it through an optical character recognition (OCR) tool.

Document analysis in Microsoft Sustainability Manager

Another scenario where Copilot clearly works with documents is the document analysis in Microsoft Sustainability ManagerThis functionality allows you to upload between one and five documents related to ESG (environmental, social and governance) and then ask Copilot questions about their content.

The flow is simple: from the Analysis section you choose “Document Analysis (Preliminary Version)”You upload the desired files, and they are securely stored in Dataverse. The system then enters a processing phase that can take up to 30 minutes; once it's ready, you receive a notification indicating that you can begin querying the model.

Then you select which documents you want to use as a reference (up to five simultaneously), open the Copilot panel and formulate your questionsCopilot only analyzes the marked documents and searches for the most relevant answer based on those sources.

A typical example is uploading a technical guide in PDF format, for example on calculating scope 3 emissions, and asking questions like “What are Scope 3 emissions and why are they important?”Copilot asks questions such as, “What methods does the guide propose for calculating them?” or “What role does a specific entity play in the development of the document?” Copilot responds by referencing the relevant sections of the uploaded files.

Again, the key point is that these documents, even if they are in PDF format, must be text-readableIf the PDFs are scanned without a text layer, the ingestion process will need to apply OCR internally, or pre-processed versions will need to be uploaded. Without this layer, Copilot's semantic analysis cannot function correctly.

Copilot usage ideas, privacy, and reliability

Microsoft offers numerous ideas for experimenting with Copilot in Word: generating internal articles about well-being, create business plans For new services, you can write marketing plans or cover letters. You can also request recipe lists, sample table formats, or even short creative pieces like haikus.

However, the company insists that, by relying on OpenAI's GPT models, Copilot can produce unexpected results or imprecise. It is important to review the content, validate sensitive data, and keep in mind that similar requests can generate very similar texts between different users.

Regarding privacy and social permissionCopilot relies on the security, compliance and data protection framework from Microsoft 365. Documents are respected according to the organization's permissions and chats are saved as part of the experience, so you can pick them up again later, always within the limits and policies of your tenant.

Microsoft actively encourages users to Send feedback Using the thumbs up or down buttons on Copilot responses. This feedback helps refine the quality of responses and improve the experience in products like Word, Sustainability Manager, and Copilot Studio itself.

Regarding general limitations, Microsoft details that some features are only available with certain licenses, that document length influences the quality of summaries, and that, although the models are powerful, They do not completely replace human reviewespecially in legal, financial or regulatory compliance contexts.

Given all of the above, the key to working with Copilot and scanned documents It's about accepting that the model doesn't magically "see" the text within an image: you need to integrate OCR (for example, with Azure AI Document Intelligence), generate a reliable text layer, and then leverage all of Copilot's capabilities in Studio, Word, and Sustainability Manager to summarize, rewrite, extract data, and answer complex questions quickly and consistently.