Catalog / Everyday Automation & AI Micro-Tools / AUTO-005

Document version comparison & change summaries

Employees compare two versions of an offer, specification, policy or terms document and get an exact change list plus a plain-language explanation of what changed and what it means.

Quick win: proven, manageable effort, first results within weeks, low regulatory risk.

Parts needed to run it yourselfAgents & automationDocument readingLanguage model

A supplier sends "version 3" of the offer, a colleague revises the specification, the policy is updated again, and nobody says what changed. Word and Acrobat can mark every difference, but a 40-page redline is hard to read. Combining the exact comparison with an AI summary gives a short list: what changed, where, and which changes matter. Typical benefit: minutes instead of an hour of reading side by side, and fewer overlooked changes in prices, dates and obligations.

Problem & value

  • Who: sales admin, purchasing, project and quality managers, office managers and anyone who receives revised offers, specifications, internal policies, general terms or tender documents.
  • Today: two windows side by side, or "please tell me what you changed" by email. Silent changes to a delivery date, a payment term or a liability sentence slip through.
  • Why AI helps: classical comparison is exact but noisy (formatting, moved paragraphs). A language model explains the differences in words, groups them by topic and flags those that look commercially or legally relevant. The exact diff remains the evidence; the AI text is the reading aid.

Variants

  • Offer and quotation revisions (purchasing, sales): price, quantity, delivery, validity.
  • Specifications and requirement documents: changed requirements for engineering and project teams.
  • Out of scope: clause-level legal redlining and negotiation (DOC-010) and source code diffs (CODE-003).

How it works

  1. two versions (docx / PDF / scan)
  2. text extraction (OCR if scanned)
  3. exact comparison (Word Compare, Acrobat Compare, or a text diff)
  4. diff + relevant context passed to an LLM
  5. grouped summary, "needs decision" list
  6. human reads summary AND spot-checks the marked diff

The deterministic diff is the source of truth; the model only explains it, because LLMs can overlook small changes. Human in the loop: the person who decides on the document always checks numbers, dates and obligations in the diff view.

Data preparation

  • Data: the two files, typically on file shares, SharePoint/OneDrive, Nextcloud, in the DMS or as email attachments. No training data is needed.
  • Steps (about 1-3 hours once): agree on a naming scheme with version numbers; keep Word originals, not only PDFs; run OCR on scans (Acrobat, or Paperless-ngx with OCR if you already use it); remove or mask personal data that the summary does not need.
  • Effort share: low. See GUIDE-003 for the general readiness check and DOC-014 for PDF-to-text conversion.

Implementation paths

Which path for whom

  • SMB: A. Word Compare, Acrobat Compare and Copilot/Gemini are already licensed or cheap; the process is standard.
  • Enterprise: B. Embed the comparison into document workflows (approval, contract repository) via API.
  • Developer: C. A small script with a text diff and a local model; no data leaves the building.
  • Private: A. Built-in compare functions and a chat assistant, with privacy settings checked.

Path A - Buy (off-the-shelf)

  • Exact diff: Microsoft Word, Review > Compare, produces a third document with revision marks and leaves the originals unchanged [1]. Adobe Acrobat Pro "Compare files" reports differences in text, fonts, images and page order for PDFs [2].
  • AI summary: Microsoft 365 Copilot can compare files from OneDrive/SharePoint with a prompt and list differences; community and training sources describe this as a summary rather than an exact diff [3]. Gemini in Google Workspace and Nextcloud Assistant (with a local or EU-hosted model) are alternatives.
  • EU options: Microsoft 365 with EU Data Boundary and a signed AVV; Nextcloud Hub on EU hosting or own server; Mistral Le Chat (French provider).

Path B - Build on cloud APIs

  • Stack: n8n or Power Automate watches a folder or mailbox, converts both versions to text, computes a diff (Python difflib or a diff node) and sends only the diff with some context to an EU-region API (Azure OpenAI in an EU region, Mistral API, IONOS AI Model Hub or STACKIT AI Model Serving).
  • Effort: one to three days for a first flow; keep the prompt and a test set under version control.

Path C - Self-host (open source)

  • Stack: n8n or a Python script, Paperless-ngx or Tesseract for OCR, a text diff, and Ollama with a small open-weight model (e.g. Qwen or Mistral class, 7-14B quantised) for the summary.
  • Hardware: HW1 is enough for short diffs on a laptop or CPU server; HW2 (one 24 GB GPU) is more comfortable for long documents and several users.
  • When: very sensitive material such as HR policies, trade secrets, health data, or when the same server already runs other AUTO and DOC use cases (Baukasten effect). Ops effort: a few hours a month.

Path D - Custom training

Not worthwhile; a good prompt template is enough.

State of the art (2026)

  • The reliable pattern is deterministic diff first, LLM explanation second. Asking a model to find differences between two long texts directly is less reliable than giving it a computed diff.
  • Limits and failure modes: missed small numeric changes, hallucinated changes, loss of tables and footnotes in PDF text extraction, moved paragraphs reported as deleted plus added, scanned documents with OCR errors producing false differences, and no judgement about legal consequences.

Shared resources & combinations

  • Same OCR and conversion stack as DOC-014.
  • Escalate to DOC-010 for legal review and DOC-016 for rule-based compliance checks.

Effort, cost & KPIs

  • Path A: hours to set up a prompt template and a short instruction; cost is existing licences, or Copilot seats where the organisation already buys them (check current list prices with your reseller).
  • Path B: 1-3 days for the flow, plus API usage that is small for short diffs.
  • Path C: 2-5 days including OCR and tests; hardware you may already own.
  • KPIs: minutes per comparison (before/after), share of changes found in a seeded test (plant 5 known changes), user satisfaction, number of "missed change" incidents.

People & change

Involve early: the department lead who owns the document type, end users, IT (licences, folder access and permissions) and the DPO. Management sets the rule which documents may go to which tool (GUIDE-007).

Works council: depends. An isolated compare function without user logs rarely triggers co-determination. Microsoft 365 Copilot and similar suites log usage and are objectively suited to monitor behaviour or performance, which the BAG treats as sufficient for §87(1) no. 6 BetrVG. §90(1) no. 3 requires timely information when work processes are planned with AI. Public bodies involve the Personalrat. Only companies with an elected works council are affected; a short works agreement on log access settles it (GUIDE-004).

Data protection: offers and contracts contain names and contact data; sign an AVV (Art. 28 GDPR) with each provider, prefer EU hosting, and check whether a DPIA is needed when HR documents or special categories are compared at scale (GUIDE-005, GUIDE-010).

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy. A 45-minute session is enough: how it works, why the diff is the evidence, typical misses (GUIDE-006).

Pilot: 5-10 people in purchasing and sales admin for 3-4 weeks with an AI champion; no sign-off on the AI summary alone (GUIDE-008, GUIDE-009).

Risks, regulation & compliance

  • EU AI Act: minimal risk; no Annex III use. If the comparison is used to evaluate employees or applicants (e.g. comparing application documents) the case changes, see GUIDE-014. Art. 50 transparency is not triggered by internal summaries.
  • GDPR: personal data in documents; AVV, EU hosting, no personal data in free consumer tools (GUIDE-005).
  • Risk of overreliance: a missed change in a liability clause. Mitigation: show the full diff, seed tests, require sign-off by the document owner.

Getting started

  1. Pick one recurring case (e.g. supplier offer revisions) and collect three real version pairs.
  2. Compare them with Word Compare or Acrobat, and write down how long it takes today.
  3. Test Copilot, Gemini or a local model with a fixed prompt: "Summarise the changes by topic; list changes in prices, dates, quantities and obligations separately; quote the old and new wording."
  4. Plant five known changes in a test pair and check what the tool finds.
  5. Document the rule in the AI policy: which documents, which tool, who checks (GUIDE-001, GUIDE-009).

Sources

  1. Microsoft Support, "Compare and merge two versions of a document" - https://support.microsoft.com/en-us/word/compare-and-merge-two-versions-of-a-document (accessed 2026-10)
  2. Adobe, "Compare two versions of a PDF file (Acrobat Pro)" - https://helpx.adobe.com/acrobat/using/compare-documents.html (accessed 2026-10)
  3. Microsoft Adoption, "Finance scenario: Amendment comparisons" (Copilot scenario library) - https://adoption.microsoft.com/en-us/copilot-scenario-library/finance/amendment-comparisons (accessed 2026-10)
  4. Microsoft Support, "Compare document differences using the legal blackline option" - https://support.microsoft.com/en-US/Word/compare-document-differences-using-the-legal-blackline-option (accessed 2026-10)