Which tool to choose
For texts pasted by hand (a long email, an article, meeting minutes) the latest free version of any general-purpose AI assistant is fine: the difference is made by the prompt, not the tool.
When the text is a whole document — a thirty-page PDF, a contract, a handout — it's better to use an assistant that reads the attached file without you pasting it in pieces. The AIs that show the source of each extracted sentence are the most suitable here: they let you trace back to the exact line they took the information from, so you verify instead of trusting.
How to do it
Extraction succeeds when you give the AI three things: the text, the precise criterion of what to keep, the form of the result. Skipping one is the reason a vague summary comes back.
- Open a new chat. A chat already full of other things confuses the extraction criterion.
- Paste the text (or attach the document). If you exceed the box's character limit, split it into two or three parts and announce it: "I'll send you the text in three parts, wait until I finish before replying".
- Write what you want to extract, naming the exact categories. Not "the important things", but "the deadlines, the amounts, and the names of those responsible".
- Indicate the form: bullet list, table, short sentences. The form forces the AI to discard the superfluous.
- Ask it to leave a placeholder where the information is missing, instead of inventing it.
The operational syntax for meeting minutes:
From these minutes extract only:
- the approved decisions
- for each decision, who carries it out and by when
Return a table with three columns: Decision, Person responsible, Deadline.
If a deadline is not indicated in the text, write "not indicated". Don't add decisions that aren't written.
After the answer, do a spot check: take a row of the table and look for it in the original text. If you find it, the rest is reliable; if you don't find it, the AI is inventing and needs to be corrected.
A concrete example
Marco receives the condominium regulations, twenty-two pages, and only needs to know what changes for those who keep a dog. He pastes the document and writes: "From these regulations extract only the articles that concern pets. For each one, report the article number and what it requires in one sentence". The AI returns four articles with the number. Marco opens the PDF, goes to the cited article 14, checks: it matches. In two minutes he has what he needed without reading the other twenty pages.
When it does NOT work (and how to fix it)
If the AI returns a summary instead of the extraction
You asked "summarize" or "tell me the important things": these are commands that produce prose, not extraction. Rephrase with a verb that isolates ("extract", "list", "find all occurrences of") and impose the list or table form. The rigid form is what blocks the tendency to summarize.
If it invents details that aren't in the text
It happens more when the text is long and the question vague. Add the constraint to the prompt "use exclusively what is written in the text; if a piece of information isn't there, write that it's missing". Then check a sample: an extracted row must be findable in the original.
If the text is too long and the AI seems to forget the beginning
Beyond a certain length the AI keeps the final parts of the text it received in better focus. Split the document into thematic blocks, have it extract one block at a time in separate messages, and at the end ask it to combine the results. Working in blocks keeps the precision that a single, giant text makes you lose.
If you can't find the button to attach a file
The names and positions of the buttons change often. If you don't see the paperclip or the upload icon, copy the text and paste it directly into the box, or write to the AI: "Act as an information extractor: I'll paste the text in blocks and you wait for all the blocks before replying".
A tip from someone who really uses it
Build your extraction prompt once for the type of document you handle most often — minutes, invoices, contracts — and save it in a note on your phone. Each time you only change the pasted text. The part that saves time is not writing the brilliant prompt every time, it's not having to rewrite it.
Frequently asked questions
How much text can I paste at once?
It depends on the tool's box: some accept a few thousand words, others whole documents via attachment. When you exceed the limit you notice because the answer ignores the final part. The rule of thumb: if the text is as long as a chapter, attach it as a file or split it into blocks.
Can I trust the extraction without rereading the original?
For decisions that matter — money, legal deadlines, health — no. The spot check costs thirty seconds: take two extracted rows and find them again in the text. If both match, the extraction holds.
Does the AI understand a text in Italian as well as in English?
For extraction yes: finding deadlines or names in an Italian text is a task that recent AIs handle without problems. The language of the text is not the obstacle; the vagueness of the request is.
Isn't it faster to use the computer's "find word" search?
Text search finds the exact words you already know. Extraction with AI finds the concepts even when they're written with different words: you look for "the deadlines" and it also picks up "by the end of the month" and "no later than the 30th", which a word-finder would miss.