Documents & Data

Next to the text box in every major AI chat sits a small paperclip or plus button. Most beginners never touch it. It might be the single most useful button in the whole app — because once AI can read your document, its answers stop being general and start being about your actual life.

The attach button is a superpower

Click the attach button, choose a PDF, Word file, or photo, and the AI reads it — the whole thing, in seconds. Then you ask questions about it in plain language. What that unlocks:

  • Summarise. "Summarise this 30-page report in one page, then in 5 bullets."
  • Explain the scary clause. Upload the rental agreement: "What does clause 7 actually mean for me? What am I agreeing to, in plain words?"
  • Extract the key numbers. "From this insurance policy: what's covered, what's excluded, what's the deductible, and when does it renew?"
  • Compare two versions. Upload both: "What changed between last year's contract and this year's? List every difference, biggest first."
  • Turn notes into minutes. Paste your messy meeting notes: "Turn these into proper minutes: decisions, action items with owners, open questions."
  • Translate a scanned notice. A photographed government notice in dense official language: "Translate this into simple English, then tell me if I need to do anything and by when."
Photos count as documents

No PDF? No problem. Photograph it: the electricity bill, the prescription, the printed form, the restaurant menu in a foreign language, the error message on your screen. If a human could read it, the AI usually can too — including a surprising amount of handwriting. "Here's a photo of my bill — why is it 40% higher than last month?" is a perfectly good question.

Why answers from your file beat answers from thin air

Here's the key idea, and it's worth a diagram. When you ask AI a question from nothing — "what does a standard rental agreement say about deposits?" — it answers from its general training: an averaged memory of many documents, none of them yours. When you attach your agreement and ask the same question, the answer is grounded: tied to the actual text sitting in the conversation. Grounded answers are dramatically more reliable, because the AI is reading, not remembering.

Asking from thin air vs. asking about your file
Without your document "What does clause 7 mean?" no file attached Answers from memory "typically, clause 7 might say…" a guess about the average contract — not about YOURS With your document Your PDF + same question the file travels with the question Reads the actual text "YOUR clause 7 says the deposit…" Grounded answer quotable, checkable, about your life
Same question, different worlds. Left: a plausible guess. Right: reading comprehension on your actual file.

A power move on top: ask it to quote the document. "Answer, and for each point quote the exact sentence from the file you based it on." Now you can verify every claim in ten seconds.

Try it now 
[Attach a contract, policy, or official letter first]

You are helping me understand this document. I'm not a lawyer.
1. Summarise what this document is and what it commits me to, in plain language.
2. List the 5 clauses that matter most to me, each with: what it says (quote it), what it means in practice, and any risk to me.
3. Flag anything unusual, one-sided, or worth negotiating.
4. What questions should I ask before signing?

Small data: your pocket analyst

AI is surprisingly good with tables. Paste a small table straight from Excel or Google Sheets into the chat (copy-paste keeps the columns well enough), or attach the file where supported, and ask in plain language:

  • "Total each month. Which month was worst, and by how much?"
  • "Sort these expenses by category and show each category as a share of the total."
  • "I sell at these prices with these costs — what's my margin per product?"
  • "Describe a chart that would show this trend clearly, and what it would reveal."

And the underrated gem for non-coders: ask it to write Excel formulas for you. Formulas are a tiny programming language, which means AI is excellent at them:

Try it now 
I'm working in Excel. My sheet has: column A = date, column B = customer name, column C = amount, column D = paid/unpaid.

Write me formulas for:
1. Total of all unpaid amounts
2. Amount owed by one specific customer I name in cell F1
3. Highlighting rows older than 30 days and still unpaid (conditional formatting steps)

Explain each formula in one plain-English line so I learn something too.

The same works for Google Sheets, and for fixing formulas: paste the broken formula and the error message, get the corrected version plus an explanation of what was wrong. Many office workers say this alone repaid every minute they spent learning AI. If it whets your appetite, Building without Code goes further down this road.

Long documents and the context limit

Every chat has a working-memory limit — the context window — covered properly in Context & Memory. Modern tools handle whole books' worth of text, but two practical habits keep quality high on big files:

  • Ask per-section, not everything at once. "Summarise chapter 2 only", then chapter 3, beats "summarise this 300-page manual" — the AI's attention, like yours, thins out when spread across too much at once.
  • Chunk monster documents. If the file is too large to upload or the answers get vague, split it (or copy in one part at a time): "This is part 1 of 3. Just acknowledge it. I'll ask questions after part 3."
  • One document, one chat. Start a fresh conversation per document so answers don't get contaminated by an earlier file's contents.

Two cautions before you rely on it

Caution one: numbers can slip. When AI reads a big table — especially a photographed or scanned one — it occasionally misreads a digit, drops a row, or mis-adds a column. It will present the wrong total with the same calm confidence as the right one. So: spot-check the numbers that matter. Recompute one or two totals yourself, or ask "list the individual values you summed so I can check." For a contract's meaning, verify the quoted clause; for a bill, verify the arithmetic.

Caution two: think before you upload. Your document goes to the AI company's servers to be processed. For most everyday files that's a reasonable trade, but for sensitive ones — salary letters, medical reports, national ID, other people's personal data — apply the rules from Privacy Basics first: blank out what isn't needed, use paid/business settings that don't train on your data where available, and when in doubt, leave it out.

Grounded is not the same as guaranteed

Attaching a file makes answers far more reliable — not infallible. The AI can still misread, skip a page, or summarise a nuance away. The habit that closes the gap: for anything with money or legal weight, ask for exact quotes and check the two or three sentences that actually matter with your own eyes. Reading three sentences beats reading thirty pages — that's still a 90% saving.

A junior clerk with a highlighter

Think of document AI as a fast junior clerk: hand over a stack of papers and they return with the key points highlighted and a neat summary on top. Brilliant time-saver — but a good manager still glances at the highlighted lines before signing anything. Speed from the clerk, judgement from you.

Key takeaways

  • The attach button turns general AI into your AI: summaries, clause explanations, key numbers, version comparisons — about your actual files.
  • Photos are documents too: bills, prescriptions, forms, menus, error messages.
  • Grounded answers (from your file) beat thin-air answers (from memory). Ask for exact quotes so you can verify fast.
  • Paste small tables for instant analysis, and let AI write and fix your Excel formulas — a non-coder superpower.
  • For long files: ask per-section, chunk huge ones, one document per chat.
  • Spot-check the numbers, and apply privacy rules before uploading anything sensitive.