A client sends a 60-page PDF and says, “Please look at this before tomorrow.” Asking AI for a summary produces smooth prose, but the key condition may sit in a table, the timeline in an appendix, and a limitation may be written as a requirement rather than a suggestion.
A PDF is not naturally summarisation-friendly: it can mix prose, scans, tables, captions, appendices, and version history. A better workflow starts with the decision you need to make, then requires the model to preserve source locations for every answer. It works for proposals, supplier materials, manuals, research reports, and customer requirements.
Write five questions before uploading
For a supplier review, ask: what is in scope; what dates are explicit; what must the customer provide; where are price assumptions; and which risks, exceptions, and open questions need discussion? “Summarise this” spreads attention evenly and misses decisive details. Also list fields AI must not guess: money, dates, version, names, integrations, certifications, and responsibility limits. OCR scans first, then sample-check numbers, table headings, and pages.
Require an output that returns to the source
You are a document-checking assistant. Using only this PDF text, answer: scope, dates, prerequisites, pricing assumptions, customer responsibilities, risks and exceptions, and questions to confirm. For every item provide a source excerpt and page or section number. If evidence is absent, write “not stated in the document.” Do not infer or add general knowledge. Separate facts, author recommendations, and open questions.
This produces a shorter but more useful brief. When using ChatGPT or Claude, check whether the file contains personal information, confidential pricing, or protected customer material. Use an approved workspace and redact when needed.
Treat tables and appendices separately
“See the fee schedule” and “parameters are in Appendix Two” are easy to miss. List the main document, tables, and appendices separately. Ask for column names, units, scope, and conflicts with the prose. A number without its heading and footnote is not meaningful evidence.
Turn the brief into a meeting action table
The output should contain item, source evidence, owner, question to confirm, and deadline. “Supports SSO” still needs protocol, implementation responsibility, and timing. “Two-week delivery” needs a start condition and acceptance definition. AI can propose questions; the owner decides what enters the agenda.
Three-minute check
- Can every important conclusion return to a page, section, or table?
- Were money, dates, quantities, and units sample-checked?
- Did “not stated” accidentally become a positive answer?
- Are appendices, footnotes, links, and version information included?
- Does every action have an owner and concrete question?
AI is valuable for making long documents searchable and discussable. It does not take responsibility for the document. With questions and source evidence first, tomorrow’s meeting will not depend on “I think the PDF said something like that.”