Workbranch › Guides

Guide

How to compare several documents or proposals with AI

Three vendor quotes, four contract drafts or a stack of grant proposals: an AI can line them up in minutes, but only if it reads every file and doesn't fill gaps with guesses. This guide covers how many files each tool takes, a prompt for a comparison table, how to check it, and how to get to a decision.

Updated 13 November 2026

Quick answer: upload all the documents to one chat (Claude takes up to 20 files per chat, Gemini up to 10 per prompt), give the AI your own list of criteria, and ask it to extract word-for-word quotes for each criterion before it builds a side-by-side table. Then spot-check the quotes against the originals, give each criterion a weight, and ask for the trade-offs, not just a winner.

Limits and menu names below were checked against OpenAI's, Anthropic's and Google's help pages as of November 2026.

Step 1: Check how many files your tool takes

ToolFiles at onceSize per file
ChatGPTVaries by plan, model and remaining upload allowance; up to 80 uploads every 3 hours, and Free accounts get 3 uploads a dayDocuments up to 512 MB (and 2 million tokens of text); spreadsheets about 50 MB
ClaudeUp to 20 files per chatUp to 500 MB; PDFs up to 1,000 pages
GeminiUp to 10 files per promptUp to 100 MB (except video)

Two details matter for comparisons. Claude reads charts and images in PDFs of up to 100 pages; in longer PDFs it reads the text only. And ChatGPT may not pull exact figures reliably from scanned pages or tables saved as images, so ask senders for digital files or a spreadsheet of the prices.

Step 2: Prepare the files

  1. Rename each file after the sender or option, such as Vendor-A-quote.pdf. The AI will use these names in the table, which makes it easy to trace each column.
  2. Write your criteria first, before you read the AI's take: price, what's included, timeline, payment terms, liability, support, whatever decides it for you. Your list keeps the AI from comparing only the things that are easy to find.
  3. Upload them. In ChatGPT, select + in the prompt area, then Add photos & files. In Claude, click + in the lower left of the chat box and choose Add files or photos. In Gemini on the web, click Add files, then Upload files.

If you compare documents of the same kind often, such as quotes for every new contract, keep the criteria and your scoring rules in a ChatGPT or Claude project so each round starts from the same setup.

Step 3: Ask for quotes first, then the table

Anthropic's guidance on reducing hallucinations recommends asking for exact quotes before the analysis, letting the model say when something is missing, and limiting it to the documents you provided. Built into one prompt:

"I've attached [number] proposals for [project]. Use only these files, not general knowledge. For each of these criteria: [your list], extract the exact sentence or figure from each file, with the file name and page or section. If a file doesn't address a criterion, write NOT STATED; don't infer it. Then build a table with one row per criterion and one column per file, using short summaries of those quotes. Below the table, list the three biggest differences and anything that looks inconsistent within a single file."

Why it works:

Before you rely on the totals, ask for the numbers in a second step: "Put every price and fee in a separate table with the currency and whether it includes tax, then calculate the total cost over [period] for each option and show the calculation." In ChatGPT, data analysis can run these calculations in Python; OpenAI advises reviewing the code and assumptions it used.

Step 4: Check the table against the source

OpenAI's help center is direct about it: models can produce made-up quotes and sound sure while being wrong, so verify quotes, figures and references. A quick routine:

  1. Search each document for the quoted words (Ctrl+F or ⌘F in your PDF viewer). Do this for every price and every legal term, and a sample of the rest.
  2. Re-check every NOT STATED cell by skimming the relevant section yourself. Missed clauses are as common as invented ones.
  3. Ask the AI to audit itself: "For each cell in the table, find the supporting quote again. If you can't, mark the cell as unverified."
  4. Correct mistakes in the chat and ask for an updated table, so later answers build on the fixed version.

Step 5: Turn the comparison into a decision

  1. Weight your criteria: "Score each option from 1 to 5 per criterion, using only the verified table. Weights: price 40%, timeline 25%, support 20%, terms 15%. Show the scores and explain each one in a sentence."
  2. Test the result: "What would have to be true for the second-ranked option to be the better choice?" This shows how close the decision really is.
  3. List the questions to send back: ask for a short list of questions for each sender based on the NOT STATED cells and the inconsistencies.
  4. Save the table. In Gemini, click Export to Sheets at the bottom right of a table to save it as a new spreadsheet in Google Drive. Claude can create an Excel file you download from the chat when Code execution and file creation is on in Settings → Capabilities (on by default).

For a single long file rather than several, see how to summarize a long PDF with AI.

Good to know

If it doesn't work

Keep the follow-ups from getting lost

A comparison usually ends with questions to send and a decision date. In Workbranch, answers come from Google Gemini and each point becomes a checklist item with progress, so the questions for each sender can be ticked off as replies arrive, and a flag with a due date puts the decision on the calendar. There's a 3-day free trial.

Sources

Frequently asked questions

Is it better to compare everything in one chat or one document per chat?

One chat, so the AI sees every file at once and uses the same criteria for all of them. If the files are too large together, first ask for a one-page extract of each in separate chats, using the same criteria list, then paste the extracts into a fresh chat for the comparison.

Can the AI tell me which proposal to pick?

It can rank options against the weights you give it and point out trade-offs, but it doesn't know your budget pressure, your history with a supplier or your appetite for risk. Treat its ranking as a starting point and make the final call yourself.

What if one document is a scanned PDF?

Scans and image-based tables are where exact numbers most often go wrong. OpenAI's help pages recommend a text-based file or spreadsheet when exact values matter, so get a digital copy from the sender or retype the key figures.