Quick answer: Yes — an AI assistant can read a contract, tender pack or insurance policy that would take you an evening, and pull out the parts that matter, in minutes. The honest catch: the same speed that makes this useful makes silent omission dangerous. The tool will not tell you it skipped the renewal clause on page 27; it will just give you a confident summary without it. So the working method is: decide your question first, ask for extraction with references rather than a summary, verify every date and figure against the source yourself, and know when the document still needs a professional. This article is that method, step by step.
What long-document AI can actually do — reported vs assumed
The capability claims in this area come from the vendors, so it is worth being precise about whose statement this is. Anthropic's documentation for its Claude models states that a single request can include up to 600 PDF pages on models with the largest context windows (100 pages on others), and describes context windows of up to a million tokens — roughly several hundred pages of text (Anthropic's context window documentation). Other vendors advertise similar long-document features for their assistants.
The same documentation, to its credit, also states the limitation that matters most: as the amount of text grows, accuracy and recall degrade — Anthropic calls this context rot, and notes that curating what goes into the context matters as much as how much fits. Practical translation: the longer the document, the less reliably any single pass will catch everything, and the more weight your own checking has to carry.
What cannot be outsourced at any context size:
- Deciding what matters. The tool does not know that the notice period could sink the deal and the colour of the logo will not.
- Absence detection. It is much harder for a model (or a person) to notice what a document never says than to repeat what it does say.
- Judgement on ambiguity. Whether a clause is standard practice or a red flag usually depends on your situation, not on the text.
What to extract from any long document
Whatever the document type — supplier contract, lease, insurance policy, tender specification (the caderno de encargos, if you operate in Portugal), grant or funding conditions — the categories that decide real outcomes are short:
- Dates and deadlines. Renewal and expiry dates, notice periods, response windows, start dates.
- Money. Amounts, payment timing, indexation or uplift clauses, penalties, who pays for what.
- Obligations. Every "must", "shall" or "obligatory" — who must do what, by when, at whose cost.
- Exit routes. How each party can end the agreement, with how much notice, and what leaving costs.
- Liability. Caps, exclusions, indemnities, and any insurance requirements placed on you.
- What is missing. The protections you assumed were there, but that no section actually grants.
A five-step method that survives contact with a 40-page PDF
1. Decide the question before you open the tool. "What could hurt us if we sign this for two years?" produces a different extraction than "summarise this document". A vague question gets you a bland summary; a specific one gets you a usable shortlist.
2. Ask for extraction with references, not a summary. For example: "List every payment obligation in this document. For each one, give the clause number, the exact quote, and the parties involved." Quoted, referenced extraction can be checked; a summary can only be trusted.
3. Verify every date and figure against the source — all of them, not a sample. Two reasons. First, misquotes do happen, and they cluster around exactly the details you will act on. Second, the number-format trap: English-language documents write one thousand two hundred and thirty-four as 1,234.56, while Portuguese documents write it as 1.234,56 — and if you ask the tool to "clean up" figures, it can quietly swap the convention. Instruct it to quote every figure exactly as written, then check them.
4. Interrogate sections, not the document. Scoped questions work better than open ones: "What does section 8 say about termination?" Then ask the absence question explicitly: "Which protections are commonly included in agreements like this but not mentioned anywhere in this one?" You will still verify the answer, but the question does part of the work a skim would miss.
5. Know when to stop. This method is a reading aid, not legal advice. For consequential agreements — leases, employment contracts, significant liabilities, anything with a personal guarantee — the extraction becomes the brief you hand to a solicitor (or, in Portugal, an advogado or solicitador, as appropriate). You arrive better prepared, and pay for fewer hours of their reading time.
An illustrative example
(Illustrative scenario, not a client story.) A two-person catering business is handed a 38-page services agreement by an events venue. They set their question: "what could hurt us in year two?" The extraction, with references, returns: automatic renewal unless cancelled with 90 days' notice (clause 12.3); a fuel surcharge indexed to a published rate (clause 7); liability capped at fees paid in the previous three months, with a £5 million public-liability insurance requirement (schedule 4); and the useful absence — nothing saying what happens to amounts already paid if the venue closes. Verification catches one misquote: the first extraction reported 60 days' notice; clause 12.3 says 90. One caught error is the entire case for step 3.
Before you paste: privacy and confidentiality
- Anonymise first. Replace names, addresses and staff details with "Party A"/"Party B". It protects people, and it also makes better prompts — the tool focuses on roles, not identities.
- Know where your text goes. Consumer and business tiers of AI tools can handle uploaded data differently. Check the tool's data-use terms before pasting commercially sensitive terms, and follow whatever written AI rules your business has (a practical checklist for an AI usage policy; what to consider about data security before using AI tools).
- Never paste other people's personal data — client details, employee records — into a general-purpose tool without a specific reason and a proper basis. Anonymisation is usually both the safer and the more useful option.
Build this skill on your own documents
The method above is learnable in an afternoon on one real document — and it works best the first time on your paper, not a sample. A Personal AI Session is built for exactly that: one 60-minute online session, one-to-one, on one prepared, anonymised task — such as running one of your own live contracts through this method — for £75 per session (89 € in Portugal), with a short written recap within two working days and one clarification email thread, up to 15 minutes in total, within seven days of the recap. Sessions are available to request now; the page explains what is included and how to request a slot.
If you would rather first test whether any of this fits your situation, book a free 30-minute call — no preparation needed, and if a session would not be the right spend for now, we will say so plainly.
Questions business owners ask about this
Is it safe to upload a contract to an AI tool? It depends on what is in it and what the tool does with inputs. Anonymise first, check the data-use terms for your tier of service, and keep a written rule about what may and may not be pasted. For most small businesses the practical risk is not hacking; it is routine, unthinking pasting of confidential or personal data.
Which AI tool is best for long documents? Several handle long documents well, and vendors' documented limits change often — Anthropic's figures above are one documented example, not a recommendation of any product. The extract-and-verify method works in any capable assistant; the tool matters less than the checking habit.
Can this replace a solicitor? No. It replaces the unread document, not the professional judgement. Use it to arrive at the review prepared: extraction verified, questions written down, costs bounded.
Related reading: checking AI translations before you rely on them and checking AI-written spreadsheet formulas — the same verify-before-trust discipline, in two other everyday tasks.
Last updated: 10 November 2026