How to Upload a Document to an AI Assistant (2026)

You are reading Chapter 7 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

Uploading a document to an AI assistant takes one click. Paste short text straight into the message box, or attach the file with the paperclip at the left edge of that box. The part worth learning is which route to pick, and how to tell what the assistant actually read once your file is in.

This guide covers the difference between an answer built from your own material and one built from general knowledge, when to paste and when to attach, what each of the four assistants accepts, and a 30-second check that catches what it skipped.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

Up to now your assistant has answered every question from general knowledge. It has read an enormous amount about how contracts, board papers and reorganisations usually work, and nothing at all about yours.

That gap is why so many answers come back sounding competent and useless at the same time. Closing it takes one click, and it is the biggest single jump in usefulness in this track.

This chapter assumes Chapters 4 to 6. If you have skipped ahead, Chapter 6 is the one immediately before this, and you need three things here: your open account, the habit of writing a clear instruction, and one real document from your own work sitting somewhere you can find it. A contract, a report, a deck, a spreadsheet, or the screenshot you took last week.

Take a supplier renewal that lands on your desk. Ask, with nothing attached, what you should push back on. You get a good answer: check the auto-renewal, look at the notice period, watch the price uplift, confirm the support terms. Every line is true, and about contracts in general rather than about yours.

Now attach the contract and ask the identical question. The reply tells you the notice window runs to 90 days and you are 11 days inside it, that the uplift is written as inflation plus 4% rather than the flat 3% you remembered, and that a clause halfway down moves support to business hours only.

The assistant did not get cleverer between those two answers. It stopped guessing. Attaching a file changes what it can see, not how well it thinks.

It describes the category rather than your case, and hedges, because it cannot see the specifics. That is instant, needs no setup, and is genuinely useful when you want the general shape of a thing. Any specific it offers is a guess dressed as a fact.

It names clauses, figures and dates from your file, quotes the parts it used, and contradicts you when the file does. You can check every claim against the source. The cost is that you have to have the file and be allowed to upload it, and on a long one it may read only part.

One rule covers the whole chapter: if the answer depends on a document, the document goes in the chat. It costs one click, it works on all four assistants, and it works on the free tier of all of them.

What is and is not safe to upload at your company is a real question with a real answer, and Chapter 9 gives it. For this chapter, practise on something you would be comfortable emailing to a colleague.

There are two ways to get material in, and beginners agonise over the choice. The rule is simpler than the agonising: paste what you can read on a screen, attach what already exists as a file.

Copy the text, click in the message box, paste, then type your question above or below it. It has one real advantage, which is that both you and the assistant can see exactly what is being discussed, right there in the thread. Use it for an email, a paragraph, a job advert, three bullet points from a call.

Click the paperclip at the left of the message box, or drag the file onto the box, then type your question as normal. All four assistants put the paperclip in the same place. Use it when the material is long, when the layout carries meaning such as a table or a form, or when the thing is already a file and turning it into text would lose something.

If it lives on a screen you cannot export, take a screenshot and send that. It goes in the same way an attachment does.

Paste something very long into ChatGPT and watch what happens. Past roughly ten pages of text it turns your paste into an attachment on its own and shows it as a small file above your message. That is deliberate behaviour, not an error, and it applies to free accounts as well as paid. It is also a decent hint about where the line sits.

Pasting out of a PDF or a formatted document drags stray line breaks, page numbers and running headers along with it. Usually harmless. If the reply seems confused about the structure of what you sent, that is your cue: attach the original instead of arguing with it.

Four families of material cover almost everything a working professional needs to hand over. They are not read equally well, and the differences are worth knowing before you rely on one.

Word files, plain text files, and PDFs made from a document all carry a real text layer, and every assistant reads that layer directly and accurately. If you have a choice of formats, send this one.

PowerPoint files and exported decks go in, and the words come through cleanly. What comes through less well is everything the deck communicated visually: which box pointed at which, what the shading meant, what the tiny footnote under the chart said. Send the underlying document too if there is one.

All four accept Excel and CSV, and all four can see your rows and columns and describe what is in them. Whether they genuinely calculate over those numbers, or read them as text and reason about them, depends on whether the assistant is running actual code underneath. It sometimes is. Trust it to find things in a spreadsheet, and check anything it adds up. Chapter 19 takes this apart properly.

Every one of the four reads pictures, which means a screenshot is a legitimate way to get material in. A dashboard you cannot export, an error message, a page of a printed report, a photograph of a whiteboard after a meeting. When the information you need is stuck inside a system that will not give it to you, a screenshot is the way out.

The rough limits differ by assistant. ChatGPT takes documents, spreadsheets and images, with a small number of uploads a day on free accounts and a much larger allowance per 3 hours on paid ones. Claude takes up to 20 files in one chat and PDFs up to 1,000 pages, though it only reads charts and layout in PDFs of 100 pages or fewer, and Excel files need file creation switched on in settings. Gemini takes up to 10 files per message at 100MB each, and video and audio go in too. Copilot takes up to 20 files per conversation at 50MB each: Word, Excel, PowerPoint, PDF, images, and plain text or CSV.

The failure mode with attachments is not rejection. It is a confident answer built on the part of your file that happened to get read.

When you attach something very long, the assistant frequently does not read all of it. It indexes the file, searches it for passages that look relevant to your question, and answers from what it found. That is a sensible design and it works well most of the time. Its cost is the quiet failure: your file uploads without complaint, the answer arrives sounding entirely normal, and nothing in it flags that a third of the document was never opened. Ask a 200-page report a question about page 180 and you may get a fluent answer drawn from page 12.

Scans behave the same way. A PDF that came out of a scanner, or a photograph of a page, has no text layer, so the assistant reads the picture instead. On clean printed text that works surprisingly well now. On a faded fax, a skewed photograph, handwriting in a margin, or a dense financial table it misreads characters, and it does not mark its own uncertainty. A wrong digit arrives looking exactly like a right one.

The fix costs half a minute and it is the same fix every time. Three steps, worth running on anything that matters.

You get a page range, or an admission.

You get words you can search for. This step does the real work: a quote is checkable against the file, and a summary never is.

They are there, or they are not. If the words are there, the answer has a foundation. If they are not, you have learned something far more useful than the answer.

Before: you describe a contract, a report or a deck in your own words, and what comes back is true, hedged, and about the category rather than about your case. You could have written it yourself.

After: you can put your own documents, spreadsheets and screenshots in front of your assistant, you choose between pasting and attaching without thinking about it, and you know how to tell what it actually read.

The chapter closes with 3 exercises, about 15 minutes: run the same request twice, once with nothing attached and once with the file, send a screenshot of something you cannot copy out, and run the quote check on the first answer.

Two things carry into Chapter 8: the document you attached, and the two answers you compared. Chapter 8 is about catching the assistant when it is wrong, and it uses exactly that pairing. The file is only half of it, because the question you ask about the file still decides what comes back, and the 26 principles of prompt engineering covers how to word it. If you want a one-page reference to keep beside you, start with our free AI cheat sheets, and the rest of the syllabus lives in the AI Central Library.

Paste what you can read on a screen, attach what already exists as a file. Attach when the layout carries meaning, such as a table, columns or a form. If the material lives on a screen you cannot export, take a screenshot and send that instead.

Claude takes up to 20 files in one chat. Gemini takes up to 10 files per message at 100MB each. Copilot takes up to 20 files per conversation at 50MB each. ChatGPT limits uploads by time rather than per chat: a small number a day on a free account, a much larger allowance per 3 hours on a paid one.

Not always. Past a certain size it indexes the file and searches it for passages relevant to your question rather than reading all of it, and it answers either way without saying which parts it skipped. Ask which pages it used, then ask it to quote the exact line its answer rests on.

Yes, by reading the picture, because a scan has no text layer. Clean printed pages read well. Faded, skewed, handwritten or densely tabulated ones get misread, and the assistant does not mark its own uncertainty, so a wrong digit arrives looking exactly like a right one.

Yes. All four assistants read images, so a screenshot is a legitimate way to hand over a dashboard you cannot export, an error message, or a page of a printed report. Send the original file when you have it, since a document with a real text layer is read more reliably.