How to summarize PDFs and contracts with AI

Long documents are where AI assistants save the most time and where careless use costs the most. A vendor agreement, an audit report, a grant application or a 90-page market study all share a problem: you need the important parts, and finding them takes hours.

This guide, current as of September 2026, gives you a method that works in Claude, ChatGPT, Copilot and similar assistants. It covers preparing the file, prompts that pull out what you need, and a verification routine that keeps you honest. It is written for business readers. For contracts, it is a reading aid, not legal advice, and it doesn’t replace a lawyer when the stakes are real.

Before you start: three decisions

1. May this document go into this tool? Contracts often contain confidential terms, personal data or clauses that restrict disclosure. Check your company’s rules and the tool’s data settings first. See our company AI policy template and is it safe to put company data in ChatGPT. If the answer is no, stop here.

2. Is the file readable text? Open the PDF and try to select a sentence with your cursor. If you can, it has a text layer and works well. If it is a scan (a picture of a page), the tool has to read the image, which is less reliable for small print, tables and handwriting. If you have the original Word file, use that instead.

3. Does it fit? Every assistant can hold only so much text at once (its context window, measured in tokens). Long documents may be truncated, split up or handled through a search step. If the tool warns you, or if answers to questions about the end of the document look vague, work in sections. The method below handles this.

Step 1: Get a map first

Don’t start with “summarize this.” Ask for structure, so you know what is there.

I've attached a document. Before summarizing, give me:
1. What type of document this is and who the parties or authors are.
2. A table of contents with page numbers or section numbers, as accurately as you can.
3. The five sections most likely to matter to [my role and purpose: e.g., a small business owner deciding whether to sign].
Don't summarize yet.

Compare the table of contents with the real one. If it invents sections or misses the ending, you have learned something important about how much to trust the tool on this file, and you have learned it cheaply.

Step 2: Ask for a summary built for your purpose

A summary for a lawyer, a finance lead and a project manager should all look different. Tell the tool who you are and what you’re deciding.

I'm [role], and I need to [decide whether to renew / brief my board / prepare for a call]. Summarize the document for that purpose.

Format:
- Bottom line: 3 sentences.
- Key points: up to 8 bullets. After each, add the page or section number in brackets.
- Numbers that matter: dates, amounts, percentages, deadlines, with page references.
- Things that are missing or unclear.

Only use what is in the document. If you are unsure, say so.

The page references are the point. They let you check any claim in seconds, and they discourage the assistant from summarizing from general knowledge instead of your file.

Step 3: Work in sections for long documents

For very long files, split by chapter or section and repeat the same question on each, then combine.

Here is Section 4 (pages 22 to 31). Extract: obligations of each party, deadlines, payment terms, and anything that would cost us money or limit what we can do. Use bullets with page numbers.

After all sections, paste the section summaries into a new chat:

Here are summaries of each section. Combine them into a single two-page brief for [audience], removing duplicates and flagging any conflicts between sections.

Conflicts are common in long contracts, for instance when the main body says 30 days and an appendix says 45. Asking the tool to flag them is useful.

Step 4: Extract the fields you always look for

If you review similar documents often, build a repeatable extraction prompt and a table.

From this vendor agreement, fill in this table. If an item isn't stated, write "Not stated". Add the page or section for each.

Term start and end dates | Auto-renewal? notice period | Fees and payment schedule | Price increase rights | Termination for convenience | Liability cap | Indemnification | Data protection / confidentiality | Governing law | Who owns work product

Put the results in a spreadsheet and you can compare vendors side by side. If you do that regularly, our guide to AI with Excel or AI for Google Sheets shows how to build the comparison.

Reading contracts specifically

The same tools can help you read agreements more critically. Useful prompts:

List the clauses that put obligations on us, and the clauses that put obligations on the other party. Highlight anything unusual for this type of agreement and explain why in plain language.

Which terms in this document are most one-sided? For each, give the clause number, quote the sentence, and describe the practical impact on a company our size.

Compare this contract against our standard terms below. List each difference, the clause number, and whether it favors us or them.

[paste your standard terms or checklist]

Treat the answers as questions to put to yourself and your counsel. See also can AI review contracts for the short answer. AI is good at finding where to look. It is not good at judging whether a term is acceptable in your jurisdiction, your industry or your particular deal. Laws and enforceability vary, and a single word can change a clause’s meaning.

How to check the output

This is where you earn the time savings safely.

1. Spot-check the citations. Open the PDF and go to five or six cited pages. Does the text say what the summary says? If two or more are wrong, don’t rely on it. Rerun in smaller sections.
2. Search for key terms yourself. Use the PDF’s find function (Ctrl+F or Cmd+F) for words like “terminate,” “renew,” “liability,” “indemnify,” “exclusive,” “assign” and “fees.” Read those passages directly.
3. Verify every number. Dates, dollar amounts, percentages and time periods must match the document exactly.
4. Ask what it may have missed. For example: “Which clauses in this document did you not include in the summary? List them by section.”
5. Check exclusions and definitions. Meaning often hides in definitions (“Affiliate,” “Confidential Information”) and in words like “except,” “unless” and “notwithstanding.” Ask the tool to list these and read them.
6. Look at the attachments. Exhibits, schedules and linked policies can override the main text. Confirm the tool saw them.
7. Read the signature-critical parts in full. For anything you’re signing, read termination, payment, liability and renewal yourself, whatever the summary says.

A helpful adversarial prompt:

Now argue against your own summary. What might a careful reader see in this document that your summary understated or left out?

When to stop and call a professional

• The contract is high-value, long-term or hard to exit.
• It involves employment, real estate, intellectual property, regulated industries, or personal data.
• The AI flags something you don’t understand and you can’t verify it.
• The other side’s lawyer wrote it and the stakes are asymmetric.

Use the AI’s output to prepare: a shorter list of specific questions makes a lawyer’s time cheaper and better spent.

Common mistakes

• Summarizing without a purpose. A generic summary hides what you care about.
• Trusting a summary of a scan. Errors creep in with poor image quality.
• Skipping the definitions section. The important terms are defined there.
• Uploading something you shouldn’t. Check permission first.
• Treating “no issues found” as a clean bill of health. It means the tool didn’t flag any, not that there are none.

What the sources say

File limits from Anthropic and the American Bar Association opinion on confidentiality, which applies to contract work.

• Anthropic says Claude analyzes both text and visual elements in PDFs of 100 pages or fewer.
• Anthropic says that for PDFs from 101 to 1000 pages, Claude processes text only.
• Anthropic says you can’t upload PDFs over 1000 pages.
• The American Bar Association says a client’s informed consent is required before inputting information relating to the representation into many self-learning generative AI tools.
• The National Institute of Standards and Technology defines confabulation as the production of confidently stated but erroneous or false content.
• An arXiv paper on how language models use long contexts finds that performance significantly degrades when models must access relevant information in the middle of long contexts.

Keep going

Better prompts make every step here more reliable: see how to write AI prompts that work. For the short version of this topic, read how to summarize long documents with AI. If you also want a place to keep your findings, the meeting notes guide shows a simple tracking table.

Contract review with OpenClaw and similar tools is a hands-on session in our two-day class: first-pass review of provided sample agreements, with the limits of AI review made clear. It is not legal advice.

Sources

• Anthropic Help Center, Upload files to Claude: “PDFs of 100 pages or fewer”
• Anthropic Help Center, Upload files to Claude: “For PDFs from 101 to 1000 pages, Claude processes text only”
• Anthropic Help Center, Upload files to Claude: “You can’t upload PDFs over 1000 pages”
• American Bar Association, Formal Opinion 512: Generative Artificial Intelligence Tools (July 29, 2024): “informed consent is required prior to inputting information relating to the representation”
• NIST, AI 600-1: Artificial Intelligence Risk Management Framework: Generative AI Profile: “erroneous or false content”
• arXiv, Lost in the Middle: How Language Models Use Long Contexts: “significantly degrades when models must access relevant information in the middle of long contexts”