Catalog / Data Readiness for AI / PREP-014

AI-ready authoring: rewriting documents for retrieval

Rewrite the 20-50 most important internal documents (manuals, policies, process descriptions) so that AI assistants find and quote them correctly: self-contained sections, clear headings, real tables, alt text, FAQ blocks.

Parts needed to run it yourselfAgents & automationLanguage modelSearch & RAG

An AI assistant is only as good as the text it retrieves. Documents written for printing, such as 40-page Word manuals, scanned PDFs, slide decks with diagrams, give poor answers because retrieval cuts them into fragments that lose their context. This use case changes how important content is written: short self-contained sections, descriptive headings, real tables, described images, FAQ blocks. It costs hours per document, not licences, and improves answers in every assistant you use now or later.

Problem & value

  • Today: knowledge sits in long documents with headings like "Introduction" or "Notes", sentences such as "see above" or "as described in chapter 3", tables pasted as images, process flows as screenshots and abbreviations nobody defined. Retrieval-based assistants (Copilot, Gemini, a RAG chatbot) split these documents into chunks; a chunk that says "this does not apply to them" is useless without the surrounding text.
  • AI value: the same documents, rewritten, produce fewer wrong or empty answers, make source citations checkable and reduce the need for expensive technical fixes. Humans benefit as well: the rewrite is also better reading.
  • Evidence: vendor and community guidance for Copilot readiness consistently names clear headings, short sections, alt text and removal of stale pages as the main levers; I found no independent study with hard percentages, so plan your own before/after measurement (see KPIs) [1][2].
  • Delimitation: automatic file parsing is DOC-014; accessibility remediation for compliance is DOC-026. Here, humans (assisted by AI) change the content.

Variants

  • Product and service documentation for customer-facing chatbots (FAQ style, one question per entry).
  • Process descriptions (Verfahrensanweisungen, ISO 9001 documents): steps as numbered lists, responsibilities in a table.

How it works

  1. question list (top 50 real questions)
  2. pick the 20-50 documents that answer them
  3. AI-assisted analysis (what is unclear, missing, outdated, duplicated?)
  4. rewrite in a standard structure (human edits, subject expert approves)
  5. publish in the knowledge location (SharePoint, Nextcloud, wiki) with metadata
  6. test: ask the questions to the assistant, compare answers
  7. fix and repeat
  • Rules of thumb: one topic per section of roughly 100-400 words; headings that contain the subject ("Vacation request: approval limits", not "Approval"); repeat the subject name instead of "it" or "see above"; spell out abbreviations at first use per section; tables as real text tables with header row, never as images; one sentence of alt text or a text description for each diagram or screenshot; a short summary at the top; a "Questions and answers" block at the end; owner, valid-from date and version in the first lines.
  • AI role: a general LLM (Copilot in Word, Gemini in Docs, or a chat assistant on non-sensitive text) proposes a restructured draft, extracts FAQs and describes diagrams. The author checks every statement; the LLM must not invent content.
  • Human-in-the-loop: the subject-matter owner signs off each rewritten document.

Data preparation

Which data: the documents themselves, plus the real questions (help-desk tickets, mailbox of the HR team, chat logs) that show what the assistant must answer.

Where it lives: SharePoint/OneDrive, Google Drive, file shares, Nextcloud, wiki, PDF archives, and heads of long-serving staff.

Typical gaps: contradictory versions, expired content, scanned PDFs without text, diagrams without description, undefined abbreviations, tables as pictures, documents without owner or date.

Preparation steps: 1. Collect the top 50 questions (one to two hours with the help desk and department leads) and map them to documents (PREP-002). 2. Select 20-50 documents; retire duplicates and outdated versions first (PREP-009, PREP-013). 3. Convert scans to text (PREP-006); convert legacy formats (PREP-007). 4. Rewrite with a one-page writing guideline and a Word/wiki template; add metadata (PREP-010) and a glossary (PREP-015). 5. Build a test set of questions with correct answers (PREP-018) and re-test after each batch. 6. Fix permissions before the assistant sees the content (PREP-023).

Tools: Word styles and templates, SharePoint page templates, Nextcloud Collectives or Wiki.js, Paperless-ngx for scanned originals, Excel for the question list, Docling (open source) to check how a converter sees your document structure [3].

Rough share of effort: most of the project (the preparation is the use case). Realistic effort is 2-6 hours per document including review; 30 documents are 2-4 person-weeks spread over a month. See GUIDE-003.

Implementation paths

Which path for whom

  • SMB - Path A: the assistant you already license (Microsoft 365 Copilot, Gemini in Workspace) plus Word/SharePoint features is enough; the process is standard and no integration is needed.
  • Enterprise - Path B: large document volumes justify a scripted pipeline (analysis and FAQ generation via an EU-hosted model API, writing to SharePoint or the CMS) with review workflow.
  • Paths C and D are not meaningful here: the work is editorial; self-hosting is only worth it if documents are too sensitive for any cloud assistant, then see PREP-013.

Path A - Buy (off-the-shelf)

  • Microsoft 365: Word/Copilot rewrite and summarise functions, SharePoint pages and document libraries with metadata columns, built-in Accessibility Checker for alt text and table headers. Microsoft publishes guidance on optimising SharePoint content for Copilot [1]. Data stays within the Microsoft 365 tenant (EU Data Boundary).
  • Open source / EU-hosted: Nextcloud (EU/self-hosted) with Collectives or Text for structured pages; Wiki.js or BookStack for wikis; OnlyOffice for editing.
  • Choose A when: content is mostly internal handbooks and the assistant is already licensed.

Path B - Build on cloud APIs

  • A small n8n or Power Automate flow reads a document, calls an EU-hosted model (Mistral, Azure OpenAI in an EU region, Aleph Alpha/IONOS/STACKIT-hosted models) with the writing guideline as prompt and returns a draft plus a list of gaps for the author. Always with human release.

Path C - Self-host (open source)

Not recommended as a starting point. If needed: Ollama with a small instruct model on HW1/HW2 for drafting FAQs from confidential documents, with n8n for orchestration.

Path D - Custom training

Not applicable.

State of the art (2026)

  • Retrieval quality depends strongly on how documents are split. Modern parsers such as Docling keep heading hierarchy and tables so chunks carry their heading path [3]; this only works if the headings and tables exist in the source.
  • Office assistants only see what the user may access and what is indexed; stale or oversharing content is the main reason for bad Copilot results in practice [2].
  • Limits: LLM rewrites can silently change meaning (numbers, exceptions, legal wording); reviews are mandatory. Long FAQ blocks can duplicate content and create contradictions when only one copy is updated.

Shared resources & combinations

  • Direct follow-on: PREP-013 (authoritative RAG knowledge base), PREP-010 (metadata), PREP-015 (terminology), PREP-016 (gold-standard examples), PREP-017 (review cycles and feedback from assistant errors).
  • Same tools (Word/SharePoint templates, Paperless-ngx, n8n) serve PREP-009 and PREP-010; the same question list is the seed for the test set in PREP-018.

Effort, cost & KPIs

  • Path A: 2-6 hours per document plus one to two days for guideline and templates; costs are mostly staff time, Copilot or Gemini licences may already exist.
  • Path B: add 3-5 days for the pipeline and prompt tuning.
  • KPIs: share of the 50 test questions answered correctly with the right source (before vs. after), number of "I found nothing" answers, share of documents with owner/date, number of retired duplicates, time to answer in the help desk.

People & change

Involve early: department leads (they own the content and the review), the subject experts who actually write, IT for the document platform, data owners for access rights, and management for the time budget. The DPO joins when HR or customer documents with personal data are included (GUIDE-005).

Works council: depends. Rewriting documents does not monitor staff, but the assistant that uses them (e.g. Copilot) logs usage and is objectively suited to monitor behaviour under §87(1) no. 6 BetrVG; §90(1) no. 3 requires timely information about AI-supported work. Handbooks written by named employees should not be turned into performance data. Public bodies involve the Personalrat (GUIDE-004).

AI literacy: Art. 4 AI Act requires providers and deployers to take measures that support the AI literacy of their staff; a one-hour session on prompting for rewrites and on checking AI drafts is sufficient to start (GUIDE-006).

Communication and pilot: frame it as "make our knowledge findable", not as extra paperwork. Start with one department and its ten most asked questions; name one champion per department (GUIDE-008, GUIDE-009).

Humans in control: every rewritten document has a named owner and is approved by a human.

Risks, regulation & compliance

  • EU AI Act: minimal risk; the Art. 4 duty to support AI literacy applies. If the knowledge base later powers a customer chatbot, Art. 50 transparency duties (in force since 2 Aug 2026) apply to that chatbot, not to the authoring [4]. See GUIDE-014.
  • GDPR: remove or pseudonymise personal data from examples (PREP-022); AVV with every cloud assistant or model API (Art. 28); a DPIA is normally not needed for pure handbook content (GUIDE-005).
  • Security: hidden instructions in documents can manipulate assistants (prompt injection); restrict who may edit the knowledge location (GUIDE-013).
  • Mitigations: owner and review date on every document, test set, change log, permission clean-up.

Getting started

  1. Collect the 50 most frequent staff or customer questions (one workshop, two hours).
  2. Pick the 10 documents that should answer most of them; retire duplicates.
  3. Write a one-page style guide and a Word/SharePoint template (see rules above).
  4. Rewrite the 10 documents with AI draft and expert review; publish with owner and date.
  5. Ask the 50 questions to your assistant, record the results, repeat after two weeks (GUIDE-009).

Sources

  1. Microsoft Learn - SharePoint and Copilot best practices - https://learn.microsoft.com/en-us/sharepoint/sharepoint-copilot-best-practices
  2. Withum - Preparing your SharePoint environment for Copilot: what to audit and why - https://www.withum.com/resources/preparing-your-sharepoint-environment-for-copilot-what-to-audit-and-why/
  3. Docling project (document conversion keeping headings and tables) - https://github.com/docling-project/docling
  4. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  5. Paperless-ngx documentation - https://docs.paperless-ngx.com/