AI document analysis across spreadsheets, slides and reports

September 5, 2026 · Ask Your Docs

AI document analysis uses software to extract, retrieve, summarize, compare or calculate information from files. Start with the result you need: finding an approval condition is different from totaling a spreadsheet or reading a diagram. Choose a workflow that exposes the evidence needed to check that result.

This guide uses downloadable, fictional business files. Their expected answers are authored checks, not a claim that every AI assistant passed them. As of September 2026, account capabilities and processing routes still need to be verified in the tool you use.

Choose the task before choosing an AI analyzer

Write the business question and the evidence a reviewer would accept. “Analyze this folder” leaves both undefined. “List the campaign approvals that remain pending, with the supporting slide” produces a reviewable result. If the question includes a total, specify the period, currency and included records before the calculation starts.

A workflow can combine tasks, but each result keeps its own proof requirement. A paragraph citation does not verify arithmetic, and a correct total does not establish that a commercial exception was applied correctly.

Inspect what the file actually contains

A readable-looking document may contain several kinds of evidence. A PDF can mix selectable text with scanned pages. A spreadsheet can hold formulas and their last saved results. A slide can place a date in speaker notes while showing a different date on screen. Decide which parts the question depends on before processing the file.

Preserve the original and record its version. Check a difficult page, table or slide in the source preview before trusting the whole extraction. If headers disappear or a table loses its row relationships, fix the input or use a different processing route. Asking a model to be more careful cannot restore evidence it never received.

Google’s Document AI overview describes separate OCR, extraction and classification processors. That distinction is useful when selecting a workflow: reading a scan and retrieving a passage are different capabilities. Keep a clear record of which transformation produced the material being analyzed.

Calculate spreadsheet results with explicit filters

Download the fictional sales workbook. Its six deal rows deliberately include records that should not all be added together. The expected August won revenue in USD is $20,000 against a $25,000 target: 80% attainment and a $5,000 shortfall. Those values were checked by an independent calculation; they are not a completed ChatGPT benchmark.

Use the Deals and Targets sheets. Calculate August 2026 won revenue in USD. Show the included deal rows and excluded rows with reasons. Do not convert currencies or include open deals. Show the calculation and compare it with the USD target.

Review the selected rows before reviewing the total. A coincidentally correct number can conceal the wrong selection. Ask how dates and empty cells were handled, and inspect any code used. OpenAI documents Python-backed data analysis; a system that only retrieves extracted workbook text should not be treated as an equivalent calculation engine.

Use the Excel guide for the full exercise. The Excel converter helps inspect selected values or formulas, but does not recalculate the workbook.

Summarize slides without dropping approval conditions

Download the four-slide Harbor campaign presentation. It names Priya Shah as owner and October 12, 2026 as planned launch. The $18,000 budget is planned, and the paid-ads allocation still shows finance approval pending. A summary saying the campaign is fully approved would change the meaning of the source.

Summarize the campaign in five points. Separate plans, approvals and unresolved decisions. Cite the slide evidence for each point. Do not infer a forecast, permission or completed approval from information that is absent.

The PowerPoint summary tool lets you inspect extracted text and choose whether to include speaker notes before sending selected text for AI generation. It attaches source passages, but a real passage can still be insufficient support for a generated claim. Check each statement and its conditions.

The deck also contains an image-only marker and a notes-only rehearsal date. Use the PowerPoint reading guide to test these separately. A text-only summary should not be scored as having interpreted the slide visuals.

Compare documents with a field-by-field evidence table

A comparison needs evidence from both sides. Download the Harbor approval brief and the presentation above. Ask for a table containing launch date, campaign owner, planned budget and quotation permission, with separate columns for each source. Preserve “planned,” “pending” and “required” as meaningful status words.

Classify each row as agreement, contradiction, omission or unclear. If one file omits a detail, that alone is not a contradiction. If the documents disagree, ask whether an explicit version, effective date or authority rule resolves the difference. A newer filename is a clue to investigate, not automatic permission to discard the older source.

These files provide a concrete review case: the Word brief calls the $18,000 budget approved, while the deck labels its budget planned. Both give the same amount, but neither establishes which description supersedes the other. Record the status difference for the owner. The deck also requires written permission for a customer quotation; that additional condition must survive the comparison even though the brief focuses on legal review.

Keep unrelated quantities separate. The sales workbook and campaign budget describe different fictional activities; combining them into a return-on-investment figure would invent a relationship. Mixed-file analysis is useful when the sources actually address the same question. Record unresolved differences for the document owner instead of asking the assistant to pick the most convenient answer.

Test absence and source coverage, not just fluent answers

Include questions the files cannot answer. The sales workbook does not supply a commission rate. The campaign deck does not supply a conversion forecast. A useful response identifies the missing input and stops short of inventing a value. If you want a scenario calculation, provide a clearly labelled assumption and keep it separate from source facts.

Test an exception as well as a straightforward fact. In the Harbor materials, marketing approval does not authorize reuse of an attributed customer quotation: written permission and legal approval are separate requirements. Check whether the assistant retrieves the exception, not merely the general approval wording.

For each answer, record the actual output, source, version, missing evidence and reviewer decision. The evaluation template supports this record. Keep the expected-answer sheet outside the uploaded corpus, so a successful-looking answer cannot simply retrieve the answer key. Rerun the same questions after a source update.

Choose a reusable library or a processing workflow

For repeated questions across maintained policies, a source library can reduce repeated file selection. Ask Your Docs follows an upload, organize, connect and ask-with-sources workflow. Its private workspace retrieves extracted passages; file support alone does not establish OCR, spreadsheet recalculation, or complete slide-image and speaker-note interpretation. Use a purpose-appropriate tool when a task depends on those capabilities.

For recurring structured extraction, define the fields, missing-value rules and review process before automating. For a one-off numerical task, preserve the workbook and the calculation. For a source-backed narrative, preserve the citations alongside the delivered report. These outputs serve different reviewers.

Check who may access the originals and derived outputs, where processing occurs, and what happens when a file is replaced or deleted. Read the document processing explanation before using a private library. Begin with harmless test files, then evaluate the workflow with the people responsible for the real sources.

Frequently asked questions