How to Summarise Long Documents With AI Without Missing Details

Tell the AI what the summary is for, make it extract before it summarises (every number, date, obligation and exception, with page references), work section by section on anything over about 50 pages, then run a second pass asking what it left out and check ten facts against the original. That catches most omissions.

Details go missing for predictable reasons. Summarising is compression, and the first things compressed away are qualifiers: "unless", "except", "provided that", "within 7 days". Numbers get rounded. Material in the middle of a very long input gets less attention than the start and end. And some tools read long PDFs as text only, so tables or charts that are images inside the file simply aren't seen. Each of these has a fix.

Follow me on Instagram@sagnikteaches

Step 1: Decide what "not missing details" means for this document

"Summarise this" gives the AI no way to know which details matter. A 60-page report has thousands of details; you care about perhaps thirty. Write down what the summary is for and which kinds of detail must survive, before you upload anything.

Connect on LinkedInSagnik Bhattacharya

This list becomes part of your prompt. It's also your checklist later: if the summary doesn't mention a notification deadline and the policy has one, you know it's incomplete without reading the whole thing again.

Subscribe on YouTube@codingliquids

Step 2: Check the file before you upload it

Three quick checks save the most common failures.

1. Is the text selectable? Open the PDF and try to highlight a sentence. If the whole page highlights as an image, it's a scan without a text layer, and some tools will read it poorly or not at all. Run it through a scanner app's text recognition first, or use a tool that handles images.
2. How long is it, and does that fit your tool? Limits vary and change, so check your tool's help pages. As of September 2026, Claude accepts PDFs up to 1,000 pages but only analyses images and layout for PDFs of 100 pages or fewer; beyond that it reads the text only. ChatGPT's help pages give around 2 million tokens of text per document file. Gemini Notebook (formerly NotebookLM) takes sources of up to 500,000 words each.
3. Is it all there? Schedules, appendices, endorsements and annexes are often separate files, and they're exactly where the conditions hide. Check the contents page against what you have.

If an important table is an image inside a long PDF, split the file so that section is under your tool's visual-analysis limit, or copy the table out separately. What ChatGPT can and can't read in PDFs and photos covers the other traps.

Step 3: Extract first, summarise second

The single biggest improvement is splitting the job in two. First the AI lists every relevant detail with its location; only then does it summarise from that list. Extraction is a narrower task than summarising, so far less gets dropped.

This is an insurance policy for my self-storage business. Before any
summary, extract into a table every:
- condition or requirement I must meet
- exclusion (what isn't covered)
- excess, limit or sum insured
- deadline or time period
- definition that changes what's covered (e.g. "unoccupied")
Columns: item | exact wording (quote) | page | section.
Include items from schedules and endorsements. Don't summarise yet.
If a section has nothing relevant, say "nothing found" for that section.

Illustrative extract of the table that comes back:

Condition | "The intruder alarm shall be set and monitored at all times when
the Premises are closed to the public." | p.31 | General Conditions 6.2
Deadline  | "notify Us of any claim within 7 days" | p.27 | Claims Conditions 2.1
Definition| "Unoccupied: any building not used for 30 consecutive days" | p.12 |
Definitions
Excess    | "Escape of water: $5,000" | p.4 | Schedule

The "nothing found" instruction matters. Without it, a section with no relevant items is just absent from the output, and you can't tell whether it was checked or skipped.

Step 4: For long documents, work in sections and keep a running list

For anything over about 50 pages, or any document where the middle matters as much as the start, don't process it in one go. Work through it section by section in the same conversation, and keep a running list.

1. Ask for the document's table of contents with page ranges.
2. Run the extraction prompt on one section at a time ("pages 1-15 only").
3. After each section, ask the AI to add the new items to the running table.
4. At the end, ask for the full table, then the summary built only from that table.

It takes longer, perhaps 30-40 minutes for a long policy, but each section gets full attention, and you can see progress. If the conversation gets very long, start a new one and paste in the running table so far.

Step 5: Run the omission pass and the negation test

Once you have a summary, ask two challenging questions. The omission pass:

Compare your summary with the full document. List anything in the document
that someone relying only on your summary might need to know but that the
summary leaves out, especially conditions, exceptions and deadlines.
Give page numbers.

And the negation test, which catches qualifiers that were smoothed away:

List every sentence in your summary that states something is covered,
allowed or required. For each, check the document for any "unless",
"except", "provided that", "subject to" or time limit that qualifies it,
and quote it.

An illustrative result: "Summary says 'customers' goods are covered up to $5,000 per unit'. Qualification on p.44: 'provided that the Customer Storage Agreement prohibits the storage of hazardous materials, cash and jewellery'." That's the difference between cover you have and cover you think you have.

Step 6: Spot-check ten facts against the source

Pick ten specific facts from the final summary (numbers, dates, deadlines and conditions, not general statements) and find each one in the original using the page number. Tally how many are exactly right. Nine or ten means the summary is reliable enough for its purpose. Fewer means go back through the section where the errors cluster. Five minutes of checking is the step people skip, and it's the one that tells you whether the other five worked.

A storage facility's 84-page insurance policy, summarised properly

An illustrative self-storage owner renews a combined business policy: 84 pages including a 4-page schedule, 9 pages of definitions, general and claims conditions, and sections for buildings, business interruption, customers' goods, liability and money. One of the site's three blocks is empty, awaiting refurbishment.

First attempt: "Summarise this policy." A tidy one-page summary: what's covered, the main limits, a general excess of $2,500. It missed four things that mattered: the alarm condition on page 31; the hazardous-goods requirement for customer agreements on page 44; the 7-day claim notification on page 27; and the definition of "unoccupied" on page 12, which restricts cover on the empty block after 30 days.

Second attempt: the method above. Section-by-section extraction produced 31 items. The omission pass found two more: an annual inspection requirement for fire extinguishers and a condition about keeping the unit register off-site. The negation test surfaced the hazardous-goods proviso on customers' goods. The spot-check of ten facts found nine right and one wrong: the summary said the excess was $2,500 for all claims, but escape of water carried $5,000.

Time and outcome. About 50 minutes in total, against perhaps three hours reading the policy closely. The owner called the broker about the empty block, arranged for the unoccupied-buildings condition to be dealt with during the refurbishment, and updated the customer agreement to name the prohibited goods explicitly. None of those actions came from the first summary.

The same clause, summarised two ways

It helps to see the difference on a single passage. The policy's customers' goods section, as the quick summary put it:

Customers' goods are covered up to $5,000 per unit against fire, theft and flood.

And as the method produced it, from the extraction table:

Customers' goods: up to $5,000 per unit (p.44) for fire, theft and flood, but theft only where there are visible signs of forced entry to the unit (p.45), and only if the customer storage agreement prohibits hazardous materials, cash and jewellery (p.44). Claims must be notified within 7 days (p.27). Excess $2,500; escape of water $5,000 (p.4).

The first version isn't false. It's just the version that leads to a declined claim when a customer's unit is emptied with a duplicated key and no forced entry. Every extra clause in the second version came from a page the owner can now turn to in seconds.

A building survey summarised for a landlord

An illustrative letting agency helps a landlord client buying a flat to let, and receives a 38-page building survey. The landlord wants to know what to fix before the first tenant moves in and what it will cost. A layered summary serves that better than a paragraph:

ONE LINE
Sound overall; two urgent items (electrics, roof flashing) before letting.

ONE PARAGRAPH
The surveyor rates the structure as sound. Two items are urgent: the
electrical installation needs testing by a qualified electrician (p.22) and
lead flashing at the chimney is failing (p.14). Four items need attention
within a year, and the surveyor recommends further investigation of damp
in the rear bedroom (p.19).

DEFECTS TABLE
Item                 | Urgency      | Surveyor's estimate | Page
Electrical testing   | Urgent       | Not estimated       | 22
Chimney flashing     | Urgent       | $900-$1,400         | 14
Rear bedroom damp    | Investigate  | Not estimated       | 19
Window seals (x3)    | Within a year| $600-$900           | 17
...
OPEN QUESTIONS FOR THE SURVEYOR
1. Is the damp likely to be condensation or penetrating?

Two details here are the ones generic summaries lose. "Further investigation recommended" items often have no cost and look minor, but they're where surprises live, so the prompt asks for them explicitly. And "Not estimated" is shown rather than left blank, so no one mistakes a missing figure for a zero.

Summarising several documents that don't line up

Comparing proposals is a summary problem with a twist: each document uses its own structure, so details that exist in one simply aren't mentioned in another. Picture a small tour operator, purely as an illustration, comparing three coach companies' proposals for a 12-day touring season, each 15-25 pages.

Summarising each separately produces three nice paragraphs that can't be compared. Extracting into one table with fixed rows works much better: day rate, driver hours per day, what happens beyond those hours, driver accommodation and meals, tolls and parking, fuel surcharges, cancellation terms, vehicle age, and "not mentioned" wherever a proposal is silent.

Row                    | Coach A        | Coach B          | Coach C
Day rate               | $1,150         | $1,020           | $1,090
Driver hours included  | 9              | "reasonable"     | 10
Overtime               | $85/hour       | not mentioned    | $70/hour
Driver accommodation   | client pays    | included         | client pays
Tolls and parking      | included       | client pays      | not mentioned
Fuel surcharge         | none           | "may apply" p.14 | capped at 5%

The cheapest day rate belongs to the proposal with the vaguest driver hours and a fuel surcharge that "may apply", which the table makes obvious and three separate summaries hide. The "not mentioned" cells become the questions for each company. For documents like these, Gemini Notebook is convenient because each answer cites its source, but any assistant works if you insist on the fixed rows.

Where the method still goes wrong

• The tool didn't read the whole file. Very large files are sometimes truncated without warning. Test it: "Quote the first sentence of the last page." If the answer isn't from the last page, split the document.
• Page numbers are off. The AI may use the PDF's page count while the document prints its own numbering, so "p.31" in one is p.29 in the other. Say which you want, and check one reference early.
• Quotes are paraphrases. Search the PDF for a distinctive phrase from two or three quotes. If they aren't word for word, ask again with "exact wording only".
• The running list drifts in a long conversation. After many sections, earlier items can quietly change or disappear. Ask for the full table every few sections and compare its row count with the last one; if it shrinks, restart with the table pasted in.
• Headings mislead. A condition placed under an unrelated heading ("General") gets skipped when the AI works by section titles. The omission pass is what catches it.

A layered format you can reuse

For most documents a small business needs to act on, this structure keeps detail while staying readable:

1. One line: the verdict or the main point.
2. One paragraph: the five or six things that matter most, with page references.
3. A table of every obligation, condition, date, amount or defect, from the extraction step.
4. Open questions: anything unclear, contradictory or missing, to ask the author or an adviser.
5. What wasn't covered: sections the AI couldn't read (scanned images, missing appendices).

Ask for this format at the end of the method, and save the prompt. The last section is the honest one: it tells whoever reads the summary where its limits are.

When the summary goes to someone else, add three lines at the top: which version of the document it covers (date or reference), who checked it and when, and the spot-check score. A summary of last year's policy that looks identical to this year's is a common way for old conditions to be trusted after they've changed. And send the summary with the document attached, never instead of it, so anyone acting on a detail can check the page it came from.

Which tool for which job

• One long document, careful reading: ChatGPT or Claude on a business plan, using the steps above in one conversation.
• Several related documents (a policy, its endorsements and the broker's letter): Gemini Notebook, which answers from the sources you add and cites the passage for each answer, making the spot-check quicker.
• A document you're editing in Word: Copilot in Word, if your plan includes it, can summarise and answer questions while the file stays in your Microsoft 365 account.
• Legal bundles, transcripts and anything confidential to clients: see summarising bundles and transcripts safely and whether chat-with-PDF tools are safe for client files before uploading.

When a summary isn't enough

A good summary tells you where to look; it doesn't replace looking when the stakes are high. Read the original sections yourself before signing a contract, making an insurance claim, relying on a condition to decide whether you're covered, or acting on a regulatory duty. Use the summary's page references to go straight to them. For contracts specifically, what AI contract review can and can't do for a small business sets out when a lawyer should read the document too.

The method above takes longer than typing "summarise this". That's the point: most of the extra time goes on the steps that find what a quick summary drops, and the details that get dropped are usually the ones with a deadline or a condition attached.

Further reads

• AI Document Processing: Turn PDFs and Scans Into Usable Data — When you need fields in a table rather than a summary.
• How to Redact Personal Data From Documents With AI Before Sharing — Remove personal data before a document goes into AI.
• How to Track Contract Renewals and Deadlines With AI — Turn the dates your summary finds into reminders.
• How to Turn Meeting Notes Into Tasks Automatically With AI — The same extraction habit, applied to meeting notes.
• Business Continuity Planning With AI: A Template for Small Firms — Use your insurance summary in a continuity plan.
• AI Contract Review for Small Firms: What It Catches and Misses — What AI catches and misses in professional contract review.
• How to Create SOPs From Screen Recordings With AI — Record a task once, talk through the why, and let AI draft the procedure. Covers Gemini video uploads, Loom, Scribe and Tango, redaction and the cold-run test.
• How to Analyse Customer Feedback Surveys With AI — A step-by-step method for turning hundreds of survey comments into counted themes you can trust, with prompts, a codebook and a checking routine.
• How to Get Your Records Ready to Sell Your Business, With AI — Get your accounts, contracts, stock and staff records buyer-ready over 12 to 24 months, using AI to index, summarise and spot the gaps.
• How Much Does a Chat-With-Your-Documents System Cost? — Build a document-chat budget that includes file preparation, access checks and the work needed to keep answers current.
• Can AI Sort and File Incoming Documents for Your Business? — When AI can sort and file your incoming paperwork, what it costs in the software you run, and the review pile that stops misfiles.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Claude help article on uploading files (PDF page and visual-analysis limits); OpenAI file uploads FAQ; Gemini Notebook help on source limits. Checked September 2026.