Chapter 21: Working With Documents, Spreadsheets, Images, and Data
From talking to working
Everything so far has mostly involved typing into a box. This chapter is about the shift that makes AI genuinely useful at work: handing it your actual material.
The difference is larger than it sounds. A model answering from training data is working from a general impression of the world. A model reading your contract, your spreadsheet, or your photograph is working from the specific thing in front of it, and that is both far more useful and far easier to verify, because you have the original.
It is also where the confidentiality questions become real. Hold that thought; it gets its own section near the end, and it deserves reading before you upload anything from work.
Documents
What works well
Extraction. Every deadline, obligation, party, amount, or defined term in a document. This is close to the ideal AI task — the answer is present in material you supplied, and you can check it.
Question answering. Ask a specific question of a long document instead of reading all of it. Does this lease say anything about subletting?
Summarizing at a chosen level. One paragraph, one page, or a section-by-section breakdown.
Comparison. Two versions of a contract, two policies, this year’s document against last year’s. What changed, and which changes actually matter? Genuinely strong, and tedious enough by hand that most people skip it.
Plain-language translation. Turning dense professional prose into something a normal person can act on.
Where it fails
Very long documents. Attention thins across a long input, and the middle suffers most. A model may summarize a two-hundred-page document competently while missing a clause on page 118. For anything critical in a long document, work section by section rather than asking about the whole.
Scanned documents and photographs of text. Character recognition errors become confident misreadings: a misread digit in a figure or date is not flagged, it is simply reported.
Tables inside PDFs. Notoriously unreliable. Columns misalign, rows merge, numbers attach to the wrong headings.
Page and section references. Frequently approximate or invented.
Anything absent. Asked what a document says about something it does not address, a model may construct a plausible answer rather than reporting the absence. Ask explicitly: if the document doesn’t address this, say so.
The rule that makes document work substantially more reliable
Require quotations.
For each item, quote the exact sentence you’re relying on and tell me roughly where it appears.
This transforms verification from re-reading the document to searching it for a phrase. A fabricated quotation is caught in seconds; a fabricated summary can survive indefinitely. Make this automatic for any document work that matters.
Spreadsheets and data
The critical distinction
One distinction governs everything here: a model can produce numbers two ways. It can write and run code that actually computes them, or it can generate what a plausible answer looks like. The outputs are visually identical. Only one is arithmetic.
Modern tools increasingly do the former, which is a genuine improvement. Do not assume it. Ask:
Did you compute that by running code, or estimate it? Show me the code.
What it does well
Description. What is in this data, what shape is it, what is the range, what looks unusual.
Cleaning. The most undervalued use in this book. Inconsistent date formats, names entered six ways, stray whitespace, mixed units, duplicates, merged categories. A tedious afternoon becomes fifteen minutes.
Anomaly spotting. Outliers, impossible values, gaps in sequences, entries that break a pattern. Excellent at drawing your eye to the twelve rows worth looking at.
Formula construction. Describe what you want in English and get the formula. Always ask for an explanation of how it works; otherwise you have a working formula you cannot maintain or debug.
Restructuring. Wide to long, splitting columns, pivoting, reshaping for a chart.
Charts and first-pass analysis. Quick visualization to see what is there.
The verification protocol for data work
Non-negotiable, and it takes about three minutes.
1. Check the row count. Did anything silently disappear during cleaning? This is the most common invisible failure.
2. Recompute two rows by hand. One ordinary, one at an extreme.
3. Check the totals. Do the parts sum to the whole? Do percentages sum to 100?
4. Sanity-check the magnitude. Is the answer the right order of magnitude for your business? You know things about your data that the model does not.
5. Look at what was excluded. Ask directly: what rows did you drop or ignore, and why?
Anything going into a decision, a filing, or a client document gets all five.
Creating files
Most tools can now produce documents, spreadsheets, slide decks, and PDFs rather than just text.
This is useful and it is a first draft, not a finished artifact. Expect to fix formatting, adjust styling, and rework layout. The content will usually be sound; the presentation usually needs a human pass. Slide decks in particular tend to arrive structurally correct and visually generic.
Two things that help. Give it a template or an example of your house format. And ask for the content and structure first, agree on those, and only then request the file: assembling a deck from an agreed outline works far better than requesting a deck outright.
Images
Reading images
Underused, and often the fastest path to an answer.
• Photograph a document, a form, a receipt, a handwritten note
• Screenshot an error message, a chart, an interface you cannot navigate
• Photograph a physical problem (a leak, a part, a plant, a rash, an appliance panel) and ask what it shows and what to check
• Upload a chart from a report and ask what it actually demonstrates, and what it obscures
The caution is the same as for scanned text: numbers read from images are a common failure point. Verify any figure that matters.
Generating images
The practical techniques are straightforward; the rights questions remain genuinely unsettled and matter the moment you move from a birthday card to anything commercial.
Before you upload anything: the part people skip
Files carry more than what you see on screen. Please read this section before uploading a work file.
A document may contain more than its visible text. Tracked changes, comments, earlier drafts in version history, author metadata, and hidden text. Pasting a client contract with the previous negotiation’s comments intact discloses considerably more than the contract.
A spreadsheet may contain more than the visible sheet. Hidden columns, hidden sheets, filtered-out rows, source data behind a summary tab, and formulas referencing other files. The tab you are looking at is often not the file you are uploading.
A photograph may contain location data, and may contain other people who did not agree to be uploaded.
Other people’s information is not yours to share. A spreadsheet of customers, a document with employee details, a thread containing someone else’s disclosures — the fact that you legitimately possess it does not mean you may put it into a third-party system.
The pre-upload check
Four questions, fifteen seconds:
1. Whose information is in here besides mine?
2. What is in this file that I am not looking at? Look for hidden rows, comments, metadata, and other tabs.
3. What are my settings? Check whether training is off, how long data is retained, and whether this is a consumer or business account.
4. What happens if this leaks? If the answer is serious, redact first or do not upload.
Redaction is usually the answer, and it is easier than people expect. Replace real names with placeholders. Strip identifiers. Upload the structure of the problem rather than the identified data. In most cases the model needs the shape of the thing, not the names attached to it.
The full “Think Before You Upload” checklist is in the Bonus Resource Library, and Chapter 25 builds the policy that governs when you use it.
Exercise 21.1: Interrogate a real document
Take a document you have signed or are about to sign. Ask for every obligation it places on you, with an exact quotation for each.
Check three of the quotations against the original. Note the hit rate.
Exercise 21.2: Clean something messy
Find a genuinely messy spreadsheet. Have it cleaned, then run all five verification steps.
Pay particular attention to the row count. This is where most people find their first silent error.
Exercise 21.3: Audit a file before uploading
Take a work file you would plausibly upload. Before you do, check it for hidden sheets, tracked changes, comments, and metadata.
Most people find something they did not know was in there.
REAL-WORLD CASE STUDY
From Raw Land to a Subdivision Application
I used AI to help complete a water-and-sewer availability application and supporting rezoning materials for a proposed residential subdivision. This was my first time handling this type of application, and I needed help understanding both the process and the technical details.
The project involved 15.64 acres and a proposed 23-lot subdivision. Translating my vision into official forms, engineering calculations, and planning terminology was daunting. Owning land does not automatically teach you how to calculate peak-hour sewer flow.
AI evaluated the subdivision concept, calculated the land devoted to the 1.61-acre right-of-way and the 0.75-acre detention pond, and helped distinguish between gross and net density. It reviewed application checklists, flagged missing information, and updated the sewer-flow calculations as the plan evolved. It also cross-checked the preliminary engineering plan against the application for consistency in such matters as water mains, elevations, and easements.
To help evaluate watershed-protection requirements, AI reviewed the applicable ordinance and estimated impervious surfaces. It converted square feet into acres and organized the figures into a usable analysis. Finally, it reviewed the completed application for missing signatures and drafted the letter of intent supporting the rezoning request.
AI did not replace the engineer, surveyor, or local officials. It helped me understand an unfamiliar process, verify calculations, and turn scattered information into a coherent submission. Most importantly, it allowed me to approach a complicated first-time assignment with far more confidence and considerably fewer headaches.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence