How to Summarise and Analyse Long PDFs and Documents With AI

Reading a 40-page contract, an annual report or a whole term’s notes takes hours. An AI assistant can give you the map in a minute and answer your questions about the text. But it needs method, because a convincing summary can leave out what matters or slip in a figure that isn’t there.

Step 1: prepare the document

• Selectable text beats a photo. Models read scans and images, but a PDF with real text gives fewer errors.
• Remove what’s not needed: irrelevant annexes, blank pages, duplicate documents.
• One document per conversation at first. Once you master the method you can compare several.
• If you’ll work with the same material often, upload it to a project.

Step 2: ask for a map, not a summary

“Before summarising, give me this document’s structure: sections, what each covers in one sentence and which pages they’re on.”

Now you know what’s inside and can steer the next questions. Then:

“Summarise the document in 7 points for someone who [role] and needs to decide [decision]. Give the page for each point.”

Step 3: specific questions, with quotes

Open questions (“what do you think?”) get vague answers. Specific ones, with quotes, get checkable answers:

“What’s the penalty if I cancel before one year? Quote the clause verbatim and give the page. If the document doesn’t say, answer ‘not stated’.”

Asking for a verbatim quote and allowing “not stated” are the two techniques that most reduce hallucinations here: they force the model to ground itself in the text instead of filling in what contracts usually say.

Step 4: extract the data to a table

“Extract into a table every deadline, amount and obligation for each party, with the page for each row.”

A table can be reviewed at a glance and pasted into a spreadsheet. If you process many similar documents (invoices, CVs), the same can be automated with structured outputs or no-code automations.

Step 5: compare versions or documents

“Compare these two contracts and make a table of the differences in price, term, cancellation and liability. Say which version favours [party].”

With very long documents, compare section by section (prices first, then terms) so the model doesn’t miss details.

Step 6: verify before using

• Open the original at the page it gave you and check every figure you’ll use.
• Be wary of round numbers or claims without a page.
• For legal or financial decisions, the summary is for understanding and preparing questions, not for replacing a professional.

Why it works (and when it fails)

Current models have huge context windows: Claude, GPT and Gemini reach around one million tokens through the API, hundreds of pages. Still, the longer the text, the easier it is to miss a detail in the middle. That’s why going section by section, asking for quotes and, for very long documents, asking per section works better.