How do you summarise PDFs and long texts reliably with AI?

What makes an AI summary reliable?

A summary is reliable if you can find every statement in the original within a few seconds. Language models write fluently, even where they are wrong; a smooth text is therefore no proof. Four qualities make the difference:

• A source reference for every statement: file, page and, where available, section or clause. For PDFs, the page number shown by the PDF program, not the printed one; Anthropic expressly recommends this for Claude.
• Figures, deadlines and names verbatim: amounts, dates and notice periods belong in the summary as quotations, not paraphrased.
• Gaps as gaps: if the document does not cover a point, the summary should say “not covered” instead of inserting a customary rule.
• One purpose, one reader: management needs different points from the people handling the case. State both in the task.

Which tools summarise PDFs and texts?

Five kinds of tool come into question. They differ in where the document is processed, how long it may be and whether the answer points to the passage in the document. The information comes from the vendors' help and product pages, as of September 2026.

Thirty contracts instead of one PDF: when does a desktop agent help?

A desktop agent helps when several documents need to be read together and the result should be stored as a file in the same folder, for example an overview of thirty maintenance contracts with term, notice period and source reference for each contract. This is how Biwak works: it reads PDFs with a text layer, Word files and emails in the selected folder and saves the summary as a Markdown or HTML file, and an overview table as CSV.

• No text recognition: Biwak reports scanned pages as images without a text layer and cannot read them.
• No audio recordings: minutes are created from notes or a transcript in text form, not from an audio file.
• No Outlook format: Biwak reads emails as .eml files, but not the Outlook format .msg.
• Data route: the documents stay in the folder, but their text goes to the language model for the summary. What this means for confidential contracts is explained in the section below.
• A single PDF in Acrobat: for this, the Acrobat Assistant with its linked source references is more convenient.

What does a good task for a summary look like?

A good task names the document, the reader and the purpose, the points that must be included, and the form of the source reference. General tips for getting started are in the first task for an AI agent.

• Basis for a decision: “Summarise the document in five points: what is it about, what needs to be decided, what does it cost, what risks does it mention, what is missing?”
• Comparing two versions: “List all substantive differences between version A and version B, with the page in both documents.”
• Deadlines and obligations: “Put all deadlines, dates and obligations into a table, with the person responsible and the source reference.”
• Short version for customers: “Explain the content in plain language on half a page, without any statement that is not in the original.”

How do you summarise long texts, emails and minutes?

Long texts are better summarised in two stages: first an outline with source references, then a summary for each section, and finally the overall picture. That way it stays visible which part fed into which statement, and no chapter drops out unnoticed. The tools' limits are generous but finite: Acrobat accepts up to 600 pages, Claude up to 1,000 pages, but Claude only analyses graphics in PDFs of up to 100 pages.

With email threads, what matters is rarely the content of each message but where things stand: what has been decided, what is open, who has to do what by when? Ask exactly that, and have the date and sender named for each point. For eligible business accounts, Copilot Chat in Outlook summarises emails directly in the mailbox; other tools need the thread as a file, for example as .eml or PDF.

Minutes are created from notes or a transcript. Have decisions, tasks with the person responsible and deadline, and open points listed separately, and give the result to the participants for approval. If you want to record meetings, you need the consent of everyone involved, because making an unauthorised recording of privately spoken words is a criminal offence under § 201 StGB (German Criminal Code). In a company, the works council may also have a say; more on this in Works councils and AI.

Where are the limits: scans, length and invented details?

• Scans: a scan is an image. Without text recognition, a tool reads nothing or guesses; OpenAI itself says that ChatGPT does not reliably extract exact values from scanned files. Make scans searchable beforehand, for example with the text recognition in Adobe Acrobat, which can also process several files in one go, and check the result.
• Length: with very long documents, some tools only analyse the text; Claude, for example, no longer analyses graphics in PDFs of more than 100 pages. Split long documents by chapter and ask whether all pages were read.
• Invented details: language models fill gaps with something plausible: a customary notice period, an amount from a similar contract, an exception that is not in the text. Such errors only show up against the original.

How do you check an AI summary?

Check a summary against this list before it is passed on. For one page, a few minutes are usually enough.

What applies to confidential documents?

Confidential documents only belong in a tool whose data route you know and with whose provider the necessary contracts are in place. If a document contains personal data, you need a data processing agreement; law firms, tax practices and medical practices must also observe professional secrecy under § 203 StGB.

Be careful with personal accounts: ChatGPT may use content from them for training unless you switch this off, and with Claude Free, Pro and Max it depends on your setting. OpenAI's and Anthropic's business plans and Copilot with a business account do not use inputs for training by default. At Biwak, processing happens in the EU, at Microsoft Azure in the EU Data Zone and as a fallback via Prem AI (contracting party PREM SA, Switzerland); for businesses, Biwak's data processing agreement applies. Even so, client and patient data does not belong in it without your own assessment.

Frequently asked questions