Summarize a PDF with AI Without Missing Pages

The 70-page tender document. The lease with 14 annexes. The annual report someone forwarded with “thoughts?” These are the files you should read and never do. AI assistants promise to boil them down to a page in seconds, and they often succeed. When they fail, you won’t notice, because a summary with holes in it reads just as smoothly as a complete one.

So this piece is less about uploading and more about one question: how do you know the summary is right?

First check: does your PDF have a text layer?

Scans are the overlooked problem. PDFs made from scanned letters often contain pictures of letters rather than text. You can test this in five seconds. Try to highlight and copy a sentence in your PDF viewer. If it works, there’s a text layer. If not, you’re looking at a pure scan.

Modern assistants can read images, but not without limits. According to Anthropic’s help centre, Claude analyses images, charts plus other graphics for PDFs of up to 100 pages. From page 101 to 1,000 it processes text only. For a scanned 300-page document, that means almost nothing after page 100 reaches the model.

Criterion	Text PDF	Scan without OCR
Body text		partly
Tables	partly	partly
Charts and diagrams	partly	partly
Long files (over 100 pages, Claude)

"Partly" depends on the tool, image quality and whether the page is analysed visually.

The fix is cheap. Run scans through optical character recognition (OCR) first. Most PDF editors have it built in, and recent versions of macOS let you select text in images anyway. After that, any assistant can read the file.

Tool limits, and the bigger limit behind them

Upload limits shift constantly, so here’s one example as of October 2026. Anthropic lists 500 MB per file, 20 files per chat and a 1,000-page ceiling per PDF for Claude chats; inside Projects the cap is 30 MB per file. ChatGPT and Gemini have their own, broadly similar limits, which you should check in their help pages before relying on them.

File size matters less than what happens next. Models reading a whole book at once don’t weigh every page equally. Researchers led by Nelson Liu showed this in the 2023 paper “Lost in the Middle”: the models they tested used information at the start and end of long inputs far better than material buried in the middle. Newer models have improved on this. The lesson still holds. With long documents, ask targeted questions rather than a bare “summarise this.”

A workflow in four prompts

Instead of one prompt, use a short sequence. It takes two extra minutes and produces something you can rely on.

1. Outline"Create a table of contents with page numbers." Shows at once whether the model read everything.
2. PurposeSay what the summary is for: a decision, a deadline, a cost.
3. Summary with evidenceEvery claim with a page number and a short verbatim quote.
4. Cross-check"What's in the document that your summary leaves out?"

Step one is a quick health check. If chapters you saw while scrolling are missing from the outline, the model didn’t capture them. Scanned pages or sheer length are the usual culprits.

For step three, this wording works well:

Summarise this document for one purpose: I need to decide whether
to give notice on this service contract before year end.

Structure:
- Term and notice period
- Fees and price increases
- Obligations that might surprise me

Rules:
- For every point, give the page number and a verbatim quote
  of no more than 20 words.
- If the document doesn't cover a point, write "not found"
  instead of guessing.

The quotes are the heart of it. Models can misread a clause; a quote with a page number takes ten seconds to verify. And allowing “not found” as an answer removes the pressure to invent something plausible. Where that pressure comes from is explained in our piece on AI hallucinations.

Contracts, official letters, research papers

Different documents need different summaries. For contracts: term, notice, fees, liability. For letters from tax offices or regulators: the deadline to respond and what you’re being asked to do. For research papers: method and sample size, not just the headline finding.

With official letters, the part that matters most often sits at the very end, in small print about how to appeal. Ask the model to quote that paragraph word for word rather than paraphrase it. Deadlines are frequently counted from the date of the letter or the date it was deemed delivered, not the day you opened it. “Roughly four weeks” is close, and close isn’t good enough for a deadline.

One point about contracts, because it’s often misunderstood. The AI can locate clauses and translate them into everyday language. It can’t judge whether a clause is enforceable in your jurisdiction or your circumstances. No summary, however good, crosses that line.

With research papers, add one more question: “What limitations do the authors themselves mention?” They usually hide in the discussion section and vanish from summaries.

Comparing two versions or quotes

One of the most useful applications barely appears in how-to guides: comparison. Two drafts of a contract, three quotes from builders, a supplier’s old and new privacy policy. By hand it’s tedious, because you keep flipping between files.

Upload both files to the same chat and ask for a table:

Compare Quote_A.pdf and Quote_B.pdf.
Build a table with these rows: total price excluding VAT,
included work, excluded work, payment terms, warranty, timeline.
Columns: Quote A | Quote B | Page in A | Page in B.
If a point is missing from one quote, write "missing".

The table often shows at a glance why one quote is cheaper. Waste removal isn’t included, or the warranty is shorter. Check the figures anyway, totals especially. Language models are unreliable at adding up amounts pulled from different places in a document. Have them list the line items and do the sum yourself.

For two contract drafts, try: “Which wording has changed? Quote the old and new versions side by side.” It’s no substitute for tracked changes, but it helps when the other side sends a fresh PDF with no markup.

Privacy and uploaded files

Contracts contain names and addresses, often bank details too. HR files even more so. Uploading them to a personal AI account means handing personal data to a vendor without a data processing agreement, which under the GDPR or UK GDPR is hard to justify.

For your own flat lease, that’s your call. Client documents are different. Either use a business plan with a proper agreement or redact personal details before upload. Our ChatGPT and GDPR guide goes through the specifics.

When a dedicated document assistant pays off

Summarising the same kind of document again and again (tenders, supplier contracts) is a sign to stop retyping the template. Custom assistants with saved instructions handle that. Our walkthrough on building your own AI assistant shows how it works in the main tools.

Thousands of internal documents that need to become searchable are another matter. Chat assistants won’t cut it; you need a system that indexes files and answers with source references. That’s the kind of build OnlineBuilders takes on for companies.

Your next step is smaller. Pick the longest unread PDF on your laptop, check whether it has a text layer, and ask for nothing but a table of contents with page numbers. You’ll find more tasks worth delegating under everyday automation.

Questions people ask

Can AI review a contract for me?

It can find clauses and explain them in plain English: notice periods, liability caps, auto-renewal terms. It can't tell you whether a clause holds up in your situation, and it carries no liability if it's wrong. For contracts with real money at stake, a lawyer still signs off.

Why does the AI ignore tables and charts in my PDF?

Depending on the tool and the page count, it may only be reading the text layer. Claude's help centre says visual elements are analysed only for PDFs up to 100 pages. Export key tables separately if they matter.

How long can the PDF be?

Limits vary by vendor and change often. Claude currently lists up to 1,000 pages and 500 MB per file in chat. Past a few hundred pages, you'll get better answers by asking about specific chapters.

Are uploaded documents private?

They sit with the vendor under its privacy terms. Documents containing client data or trade secrets belong on a business plan with a data processing agreement.