Practical Guide
How to summarise a long document with AI
The quickest way to summarise a long document with AI is to stop asking for a summary. Tell it why you are reading, ask a specific question, and make it answer with page references you can check. A request for "a summary of this report" may give you a polished tour of the whole file while the answer you need sits in a table or footnote.
You are never reading the annual report, the consultation paper or the information memorandum for its own sake. You are reading it to decide something. Say what.
Why the default summary disappoints
A general request invites broad coverage and can give too much space to recurring boilerplate: the strategy statement, the risk language, the governance section that has barely changed.
The exception that changes your decision may be mentioned once, perhaps in a middle section or a table. It is easy to lose in a short answer.
A summary that covers everything has told you nothing. You wanted a decision, not a shorter document.
How to summarise a long document with AI in four passes
Get the map first. Ask for the structure: sections, what each one covers, and which two or three are relevant to your question. Do not let it summarise anything yet. That map shows where to focus your reading.
Ask your real question, one at a time. Not "summarise the credit section" but "what security is being offered, and what happens to it on a change of control". Specific questions get specific answers because there is less room to generalise.
Ask what is odd rather than what is there. The high-value prompt in any long document is some version of: what is unusual here compared with a standard document of this type, what is missing that you would expect, and where is the language vaguest. That question surfaces the things a summary smooths over.
Then check three claims. Pick the three that would change your decision, ask for the page or section each came from, and open the document at that page. Not all of them, three. Opening those pages can catch a missing qualification before you quote it. It does not prove every other claim is right.
What gets lost
The middle is worth checking. Earlier long-context research found that some models missed facts placed there, but performance varies by model and task. A clean summary can still omit the qualification that changes your decision.
Footnotes and tables are the second casualty. In a financial or legal document, a footnote under a table can qualify the headline number. If the summary gives you the number, open the table and read the condition too.
Then there is the format problem. A scanned PDF, a chart with no underlying data, a photographed page: first check what your tool actually processed. Some tools extract text; some also inspect PDF visuals. If a claim depends on an image or table, open it yourself. The habit of checking AI for errors applies most sharply here.
One honest caveat cuts the other way: newer research finds that some models retrieve simple facts reliably even from very long inputs. The four passes are a verification method, not a permanent rule about every model. Re-test the method on your own documents.
The Hong Kong version of this problem
The documents that actually eat a week here are consultation papers from the SFC or HKMA, information memoranda, leases, and board packs that arrive on Friday for a Monday meeting. Some of those documents are confidential, so where you upload them is a real question, not a formality, and worth settling before the habit forms: see whether ChatGPT is safe for work data.
For a regulatory consultation there is a better prompt than any summary request: ask what would have to change in a firm like yours if every proposal in the paper were adopted as drafted. That gives you a starting memo to check against the proposal, including changes the model may have missed.
Where the time goes instead
None of this means you never read the document. On anything you are signing, you read the sections that matter, and a summary is a map rather than a substitute.
What changes is the ratio. Use the map to read the sections that decide it, instead of skimming every page and hoping the important qualification sticks. If you want that habit built on the documents your own desk actually receives, start here, or take the interrogation prompts from the free prompt library and try it on whatever landed this week.
Start with the question. Check the page.
Build a repeatable way to read the documents your own desk receives, with the right tool and the right checks.
Request a Private Consultation