You can trust an AI summary of a long document only to the extent that you can check it. The reliable method is: tell the AI what the summary is for, summarise section by section, ask it to quote the source for every key point, then verify the points that matter against the original. Treat the summary as a fast first pass that tells you where to read, not as a replacement for reading the parts that carry risk.
Why long documents are harder than they look
An AI assistant can read a surprising amount of text in one go, but length still causes problems. Details in the middle of a very long document can be given less weight than the opening and the conclusion. Caveats, exceptions and conditions are short sentences that a summary tends to drop, and they are often the most important lines in a contract, a policy or a report. A fluent, confident summary can hide the fact that something was left out. See our explainer on the context window for why there are limits on how much the model works with at once.
Step 1: say what the summary is for
"Summarise this" produces a generic summary. A better instruction names the reader and the decision:
- "Summarise this vendor proposal for a finance director deciding whether to shortlist it."
- "Summarise this policy for new staff: what must they do, and what must they never do?"
- "Summarise this report for a board paper: findings, risks and recommended actions only."
The purpose tells the AI what to keep and what to cut. It also gives you a way to judge the result: does it answer the decision you actually face?
Step 2: work in sections, then combine
For a long document, split it by its own structure, such as chapters, clauses or appendices, and summarise each part separately. Then ask the AI to combine the section summaries into one overview. This keeps each pass focused and makes errors easier to trace, because you know which section a claim came from. If the document is too large to paste whole, sections are also the natural way to work around that limit.
Step 3: ask for quotes and locations
The single most useful habit is to require evidence. Add to your instruction:
- "For every key point, give a short direct quote from the document and the section or page it comes from."
- "If the document does not say something, write 'not stated' rather than inferring it."
- "List any conditions, exceptions, deadlines and amounts mentioned, exactly as written."
Quotes turn checking from a search into a lookup. If a quote cannot be found in the source, you have caught a fabrication. This is the same risk described in our article on the AI hallucination.
Step 4: check against the source
You do not need to re-read everything. Match your effort to the stakes:
- Low stakes (an internal article you are skimming): spot-check two or three points.
- Medium stakes (a report that shapes a recommendation): verify every quote and every number.
- High stakes (contracts, compliance, anything with legal or financial consequences): use the summary only to navigate, and read the relevant clauses yourself. For legal documents, this is general information and not legal advice; confirm with your legal or compliance team.
Our guide on how to check whether AI output is accurate covers verification habits in more detail.
Step 5: hunt for what is missing
After the summary, ask a second question: "What important caveats, exceptions, risks or disagreements in the document are not reflected in your summary?" Then ask: "What would a sceptical reader challenge in this document?" These prompts often surface the sentences the first pass smoothed over. You can also ask for the opposite view of the document's main claim to check that the summary has not just echoed the author's framing.
Protect confidential documents
Before you paste a document into any AI tool, check whether it contains customer data, personal data or confidential commercial terms, and whether your company policy allows that tool for that material. Read can you put customer data into Claude? and your own AI usage policy first. Remove or mask sensitive details when the task does not need them.
Build the skill in your team
Summarising with evidence is a core everyday skill, and it is a habit that is easier to learn with practice on real documents than from a slide. Agmo Studio, a Select Partner of the Anthropic Claude Partner Network, runs hands-on workshops through Agmo Academy, including the Claude Certified Associate workshop. See the overview of AI training in Malaysia for programs and how HRD Corp funding may apply, subject to eligibility and HRD Corp approval.