What you’ll be able to do

You can upload a PDF to an AI assistant and just ask it things. That’s really the whole trick. A contract, an old scanned letter, a spreadsheet exported as a PDF with fifty tables in it, doesn’t matter. You upload it, you ask, you get an answer.

By the end of this you’ll be able to pull specific facts out of a document, get a summary without reading the whole thing yourself, handle a scanned PDF that’s really just a photo of text, and pull numbers out of a table without the AI mangling them (or at least know when it has). You’ll also know how to check its working, because it does get things wrong, and you need to catch that before you trust it with anything that matters.

Before you start

A couple of things worth knowing before you drag a file into a chat box.

First, privacy. Whatever you upload to ChatGPT, Claude, Gemini or Copilot is being sent to someone else’s servers and processed there. For a council tax bill or a recipe book that’s fine. For a document with your bank details, medical records, someone else’s personal data, or anything under an NDA, it’s not. I’ll come back to this properly later, but keep it in the back of your mind now.

Second, file size. Most of these tools have upload limits, and they’re not huge. A ten-page letter is nothing. A four-hundred-page manual might not go through in one piece. You’ll find out fast, because it just won’t upload.

Third, and this trips people up more than anything else: not all PDFs are the same. A normal PDF has actual text in it, the kind you can highlight and copy with your mouse. A scanned PDF is a photograph of a page, saved as a PDF. It looks like text but it’s really just an image, and the AI can’t read the words unless it either has vision (so it’s looking at the picture) or someone runs OCR on it first, which converts the picture into actual readable text.

How do you tell which one you’ve got? Try selecting a line of text with your cursor. If it highlights like normal text, you’re fine. If nothing happens, or the whole page selects like one giant image, it’s scanned. This matters a lot for what you do next.

Get set up

For most people, most of the time, the easiest route is just uploading the PDF straight into ChatGPT, Claude or Gemini and asking questions in the chat. All three let you attach a file and start talking to it. No setup, no faff, just drag it in.

If you’ve got several PDFs you want to ask questions across, say a folder of research papers or three versions of the same contract, NotebookLM (from Google) is worth knowing about. You upload multiple documents as “sources” and it answers with citations pointing back to which document and which bit it pulled the answer from. Genuinely useful if you’re working across more than one file at a time.

If you’re already inside Adobe Acrobat, there’s an AI Assistant built in that can summarise and answer questions about the PDF you’ve got open, and it’s also useful for scanned documents because Acrobat has proper OCR built in.

To actually get going: open your AI assistant of choice, click the attach or upload icon (usually a paperclip or a plus sign near the chat box), select your PDF, and wait for it to finish uploading. Then just type a question. That’s it. No accounts to create beyond the one you already use to log in, no special formatting needed on the file.

Try it yourself

Here are three prompts that each do something different. Copy them, adjust the bracketed bits, and see what comes back.

A straightforward question, once the PDF is uploaded:

Based on the PDF I've uploaded, what does it say about [topic, e.g. cancellation policy]? Quote the exact wording if you can.

Asking it to map out a long document before you dive in, which is genuinely useful for anything over about twenty pages:

Give me a section-by-section map of this document: what each section covers, roughly how long it is, and which page it starts on.

Pulling a table out cleanly, which is where most of the pain happens:

Find the table on [page number, or describe it, e.g. "the table showing monthly costs"] and reproduce it as CSV, keeping every number exactly as written in the original.

A couple of others worth having in your back pocket. Asking where something is written down forces the AI to point to a page rather than paraphrase from memory:

On which page does the document mention [term or figure]? Quote the surrounding sentence.

And asking what’s missing is oddly one of the most useful questions you can ask, because it stops the AI from just agreeing with whatever you assume is in there:

Does this document mention anything about [topic]? If not, say clearly that it isn't covered.

That last one catches a lot of mistakes early. AI assistants are generally quite happy to answer a question even when the honest answer is “that’s not in here”, so asking directly for confirmation helps.

Check the result

Don’t just take the first answer and run with it. A few quick tests.

Ask it to quote the exact sentence it’s basing an answer on. If it can produce the actual wording from the document, that’s a good sign it’s genuinely reading it rather than guessing from the general shape of the question.

If there’s a table involved, this is where I’d slow right down. Ask for the CSV version, like above, then actually check a handful of the numbers against the original PDF yourself. Tables are where AI tools go wrong most often, rows get skipped, columns shift, totals get invented. It happens more than you’d think, especially with scanned tables or oddly formatted ones. Checking three or four values against the source takes thirty seconds and can save you from a genuinely bad mistake later.

Ask the same question two different ways and see if the answers match. If you get two different figures for the same thing, that’s your signal to go and look at the page yourself.

And try the “what’s not in here” test from above on something you know isn’t in the document. If it invents an answer instead of saying “this isn’t mentioned”, treat everything else it’s told you with a bit more suspicion too.

If it doesn’t work

If the upload just fails outright, it’s almost always size. Very large PDFs, especially scanned ones with lots of images, can hit upload limits fast. The fix is usually to split the PDF into smaller chunks (there are free tools online for this, or Acrobat can do it) and upload it in parts, asking your questions section by section.

If the PDF uploads fine but the answers are rubbish, vague, or clearly wrong, check first whether it’s a scanned document. If you can’t select the text with your cursor, as I mentioned earlier, that’s your answer. Try Google Drive instead: upload the scan, right-click it, and open it with Google Docs. Google will run OCR on it automatically and convert the image into actual text, which you can then paste into any AI assistant properly. Acrobat’s built-in OCR does the same job if you’re already in that ecosystem.

If it’s not a scan issue and the answers are still off, try re-uploading. Sometimes the file just didn’t process properly the first time.

And on privacy: don’t upload anything you wouldn’t be comfortable sitting on someone else’s server. That means no passports, no bank statements, no medical files, no documents covered by an NDA or containing other people’s personal information without their consent. If you’re not sure whether something’s sensitive, treat it as if it is. There’s no version of “just this once” that’s worth the risk.

Keep exploring

If you’re mainly after a quick overview rather than detailed Q&A, have a look at “How to get AI to summarise a document”. If your scanned document is actually handwritten rather than typed, “How to get AI to read my handwriting” covers that specifically, since it behaves a bit differently. And if you’re using a PDF as one piece of a bigger research task, “How to get AI to help me research a topic” builds on a lot of what’s here.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.