How AI Reads Sales Spreadsheets, Word Specs, and PDFs
A practical guide to AI analysis of sales spreadsheets, Word specifications, PDFs, and images, including structure, citations, and limitations.
TL;DR
An AI system that can open Word, Excel, and PDF files has not necessarily understood them.
Sales documents carry meaning through structure: spreadsheet headers, merged cells and formulas; Word headings and tables; PDF layouts, scans and footnotes. Reliable analysis should preserve that structure and tell the user what was read, what may have been lost, and which numbers still need to be checked in the original file.
Office files are becoming an agent workspace
Microsoft has been moving AI from a chat sidebar into direct document work. Its announcement that agentic capabilities in Word, Excel, and PowerPoint are generally available reflects that direction. Notion has likewise expanded file types and agent workflows in Notion 3.6.
For sales teams, however, the meaningful acceptance test is not whether the agent produces a summary. It is whether it correctly locates the price, condition, unit, formula, and exception that will shape a customer response.
Three file types, three different failure modes
Excel: correct numbers can still have the wrong meaning
A quotation workbook may contain:
- several sheets;
- product, option, and service lines;
- unit price, quantity, discount, tax, and currency;
- formulas and cached display values;
- hidden rows, notes, or formatting-based cues.
If an AI extracts visible text only, it can interpret “Included” as a zero price, confuse a subtotal with a unit price, or miss the currency. A robust parser should preserve relationships among sheet names, cell ranges, formula text, and displayed values.
Word: section hierarchy changes the claim
The same capability can appear under standard configuration, optional configuration, prerequisites, or limitations. Flattening a Word document into unrelated chunks may turn an option into a standard feature.
Useful extraction keeps heading hierarchy, paragraph order, and tables. It should also be honest about layout: without a full pagination engine, a stable section or table reference is usually more reliable than an invented Word page number.
PDF and images: layout may be part of the data
Native PDFs can expose text, while scanned PDFs behave more like images. Multi-column layouts, footnotes, stamps, and diagrams can change extraction quality. Product images provide useful clues, but appearance alone should not be treated as proof of an exact configuration.
OCR and visual output should therefore remain labeled as extracted or inferred material until verified.
A synthetic workflow: preparing a response from two attachments
Customer Company A sends a spreadsheet template and a PDF of technical requirements, asking for a same-day response.
A sensible AI workflow does not immediately generate a final quote. It first:
- inventories spreadsheet sheets, currencies, key headers, and formula areas;
- extracts mandatory conditions, acceptance criteria, and exclusions from the PDF;
- compares those requirements with company product knowledge;
- labels each item as supported, possibly supported, or requiring engineering confirmation;
- prepares a gap list for the seller to confirm before drafting a response or updating account activity.
The agent helps surface risk before it helps generate content.
What KnowSales currently supports
The KnowSales Agent workspace supports questions across common sales attachments, including DOCX, XLSX, PDF, and image files. The verified focus is structural retrieval, not simply successful upload:
- Word headings, paragraphs, and tables;
- Excel sheets, cell ranges, formula text, and cached display values;
- PDF and image inputs, with result-level review still required for scans, complex layouts, and visual judgments;
- multiple attachments in the same Agent workspace, alongside company-knowledge and customer-retrieval tools, although this Office acceptance did not mark every cross-source combination as independently verified.
The boundaries matter. Legacy .doc and .xls files, macro-enabled .docm and .xlsm files, password-protected documents, and workbooks that require formula recalculation are outside the current reliable processing path. The system should not invent Word page numbers or upgrade a visual guess into a product fact.
Eight questions for evaluating document AI
| Area | What a good system should show |
|---|---|
| Formats | An explicit supported and unsupported list |
| Spreadsheet structure | Sheets, headers, ranges, units, and currencies |
| Formulas | A distinction between formula text and cached output |
| Document structure | Sections and tables, not only a full-text summary |
| Images and OCR | Extraction status and uncertainty |
| Citations | A path back to the attachment or knowledge source |
| Permissions | No retrieval by unrelated users |
| Writeback | Human confirmation before analysis becomes CRM truth |
A minimum rollout plan
Start with one repeated document type, such as the quotation workbooks your team receives every week:
- select ten de-identified samples with different headers and formulas;
- define the fields that must be extracted and the acceptable error boundary;
- require structured output with source locations;
- compare every result with the original file;
- turn recurring exceptions into explicit validation rules;
- only then connect the workflow to quotation, email, or CRM actions.
Validate “reads correctly” before optimizing “works quickly.”
Related reading
- Why AI Sales Answers Need Source Citations
- From WhatsApp and WeChat Threads to Sales Memory
- How to Build an Effective Sales Knowledge Base
Next step
If your sales knowledge is buried in quotations, specification files, and customer attachments, explore KnowSales for export sales teams.
This article is based on acceptance evidence for the KnowSales Office attachment path. The customer and document examples are synthetic.