AI for Documents and Spreadsheets That Actually Saves Hours

In Prompting for Work: Double the Output Without Learning to Code, you learned five prompt patterns that cover most office work. But when the work itself lives inside a long report, a scanned PDF or a tangled spreadsheet, the tool also has to read the document's layout and structure: turn words on a page into rows in a table, or a plain sentence into a working formula. That is what today's document and spreadsheet AI does.

Much of the time saving comes less from the model's raw intelligence than from the way these tools are built around it. They read the layout, show you a plan before they change anything, and let you correct them with a quick message instead of starting over.

How do I summarize and draft from a document?

This is the most common job, so start here. Upload the document to your company's approved AI tool and use the patterns from Part 2: a role, a clear task, a format, and a request for evidence. For example, with a 40-page supplier review:

You are an operations analyst writing for a busy director.
Summarize the attached supplier review in five bullet points: the main finding, the biggest risk, the cost impact, what changed since last year, and the recommended next step.
For each bullet, quote the sentence it comes from and give the page number.

Then check it. Open two or three of the quoted pages and confirm the quotes are really there and say what the summary claims. Tools often answer from the passages they found rather than the whole document, which How ChatGPT Reads a PDF explains. Asking for quotes and page numbers makes gaps easy to spot.

Drafting works the same way. "Using the attached review, draft a one-page email to the supplier asking about the three late deliveries, in a firm but friendly tone" gives you a first draft in seconds. Follow up with Part 2's self-critique pattern ("check this draft against the review and flag anything it gets wrong") before you send anything. And if the document is confidential, use only a tool your company has approved; AI Privacy at Work: What You Can and Can't Paste, Part 7 of this series, explains why.

How does AI actually read a scanned invoice or claim form?

Picture a careful assistant armed with a ruler and a highlighter. The assistant notes exactly where every word sits on the page, sees that "Invoice Number" is printed near the top, and after a hundred invoices learns that the number is always the bold string in the top right corner. That is the mental model behind modern document reading.

The first step is OCR (optical character recognition), which turns the image of a page into text, with the position of each word. A layout-aware model then uses where the text sits to tell a total from an invoice number. Document AI: How Models Read PDFs, Tables, and Receipts explains the models behind this in depth.

Cloud services wrap this into a ready-made pipeline. Google Cloud's Document AI, for example, has pretrained parsers such as an Invoice Parser that returns fields like invoice number, supplier name, amounts and due date, and each answer comes with a confidence score. Its OCR also reads handwriting in more than 200 languages. For your own forms, Google's custom extractor uses generative AI that can start with few or no labelled examples; Google says to expect reasonably good results with 10 to 50 or more.

Why can AI write a correct formula from a plain sentence?

Think of a spreadsheet expert who knows hundreds of functions and can write =SUMIFS(Sales[Amount], Sales[Year], 2025, Sales[Region], "Europe") the moment you say "sum 2025 sales for Europe". That expert also glances at your table headers first, so they don't refer to a column called Year23 when you meant Year.

Research showed that a language model can learn this mapping when trained on thousands of pairs of plain descriptions and their correct formulas (Zhao et al. 2024, NL2Formula: Generating Spreadsheet Formulas from Natural Language Queries). Today's tools build the same idea into the spreadsheet itself, and they look at your actual column names first, which makes them less likely to invent columns that don't exist. They can still get things wrong, so test a new formula on a few rows you can check by hand.

Copilot in Excel offers three modes: edit, plan and chat. In edit mode (labelled "Allow editing") it changes your workbook directly: cells, charts, PivotTables. In plan mode it proposes a plan before it edits anything. In chat mode ("Chat only") it answers questions and analyzes data without changing the workbook. Gemini in Google Sheets can create tables, formulas and pivot tables and format a table from a description; it needs an eligible Google Workspace or Google AI plan.

How do these tools actually save hours on a real workweek?

Let's walk through the week of Alex, our operations manager. The times are illustrative, from her fictional week.

Monday: invoice pile-up. Fifty PDF invoices land in a shared drive. Her IT team has set up a document AI pipeline that runs automatically. By the time she opens her laptop, a spreadsheet already holds vendor names, amounts, and due dates, each with a confidence score. She skims the five marked "low confidence" and corrects one typo. What used to be a two-hour typing marathon now takes about ten minutes.

Tuesday: dirty sales spreadsheet. The regional sales file has dates in three formats, duplicate rows, and a column called "Amt" that nobody can explain. Alex opens Copilot in Excel in plan mode and types "Clean this table: standardize the dates, remove duplicate rows, and rename the Amt column to Amount USD." Copilot shows its plan, she approves it, and it makes the changes. Then she asks "Show me a trend line of monthly sales by region," and Copilot builds a PivotChart on a new sheet. A job that took about an hour and a half of wrestling with functions now takes about thirty minutes, checking included.

Wednesday: report assembly. Her team keeps a shared summary in Google Sheets. Alex asks Gemini in Sheets to "create a table of the top five vendors by spend and add a formula that flags any vendor above the monthly threshold." It builds the table and the formula from the data in the sheet. What used to take an hour of copying and charting takes about twenty minutes.

The AI does not remove judgement. Alex still decides what to report and checks the output. But the mechanical parts, retyping, formatting, hunting for the right function, shrink to a fraction of their old size.

What stops the AI from making dangerous mistakes?

The guardrails live in the workflow, not just the model. Document extractors give a confidence score for each field, so in a careful setup low-confidence items go to a person instead of slipping into your ERP (enterprise resource planning) system.

Spreadsheet assistants let you control how far they go. Copilot's plan mode shows you what it intends to change before it changes it, and chat mode leaves the workbook untouched. You can also steer with quick corrections: if a formula gives the wrong result, type "That missed December" and the assistant revises it.

The same principle applies to bigger systems. A well-built setup never lets the AI post straight into your financial systems. It writes to a holding area, and entries must pass validation rules or a person's approval before they are posted. That makes mistakes easy to catch, audit and roll back.

Quick Reference

Microsoft's current name for its document service comes from its Document Intelligence overview, and ChatGPT's spreadsheet feature is described in Data analysis with ChatGPT.

Frequently Asked Questions

Q: Can AI handle scanned handwritten forms?

Often, yes. Google's Enterprise Document OCR, for example, reads handwritten text. Accuracy depends on how clear the handwriting is and whether the fields sit in predictable places. For cursive notes without a fixed layout, expect lower confidence and plan for human review.

Q: Will the AI expose sensitive financial data to a public model?

It depends on the version you use. Business and enterprise versions run under your organization's contract, and both Microsoft and Google state that they do not use that data to train their models. Personal accounts follow different rules, covered in Part 7, AI Privacy at Work: What You Can and Can't Paste. Before you upload data to any outside tool, check its data policy and your company's list of approved tools.

Q: My spreadsheet has irregular headers and merged cells. Will the AI still work?

Often better than older, rule-based tools did, because AI assistants can read the whole sheet for context. The more reliable route is still to use the AI to clean the sheet first (one header row, no merged cells), then ask for your formulas.

Q: How long does it take to set up a document extraction pipeline?

A pretrained parser, such as Google's Invoice Parser, needs no training: you send documents and get fields back. A custom extractor for your own forms takes longer, mostly to define the fields and test them on samples, but the generative AI version can start with few or no labelled examples. Connecting the results to your other systems usually takes your IT team's help.

Q: Will this replace my team's analysts?

It changes their focus more than it removes the need for them. Repetitive data entry and formula debugging fade, but judgement, handling exceptions, and explaining what the data means still need people.

Test yourself

Alex receives a batch of 200 insurance claim forms from a new partner clinic. The layout is similar to her existing templates but not identical, and the AI extractor misses the diagnosis code in about 30% of the documents. What should she do to fix this without manually reading every form?

Answer: Alex should set up a review queue for low-confidence fields. She uses the extractor's confidence scores to sort the 200 forms, and only the forms where the diagnosis code came back with low confidence, around 60 of them, go to a reviewer. The reviewer fills in the missing codes, and those corrected forms become new labelled examples for the custom extractor, which tends to improve as it gets more examples. The next batch should need far fewer reviews, then only spot checks. Because these forms hold health information, she should also make sure the extractor and the review queue are approved for that kind of data. The key is to treat the AI as part of a workflow that gets better with a handful of human decisions, not as a one-shot answer.

Next: AI for Meetings and Email: Reclaim Your Calendar. It applies the same habits to your inbox, your meeting notes and your calendar.

If this kind of concrete, no-hype breakdown of how AI fits into your workweek is your thing, subscribe to Internals Decoded at internalsdecoded.com.

Sources

• Document AI overview (Google Cloud)
• Document AI processors list, including the Invoice Parser and Enterprise Document OCR (Google Cloud)
• Handle the processing response, with confidence scores (Google Cloud)
• Custom extractor with generative AI (Google Cloud)
• NL2Formula: Generating Spreadsheet Formulas from Natural Language Queries (Zhao et al. 2024)
• Customizing outputs with Custom Queries (Amazon Textract)
• Frequently asked questions about Copilot in Excel (Microsoft)
• Collaborate with Gemini in Google Sheets (Google)
• What is Azure Document Intelligence in Foundry Tools? (Microsoft Learn)
• Data analysis with ChatGPT (OpenAI)