Catalog / Data Readiness for AI / PREP-016

Collecting gold-standard examples & templates for AI

Select, anonymise and curate the best past offers, letters, reports, replies and service reports as a small reference library that AI tools imitate, so generated drafts match company standards.

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

Parts needed to run it yourselfAgents & automationLanguage modelSearch & RAG

Generic AI drafts sound generic. Offers, letters and service reports only match your standards if the assistant sees your best past work. Pick 5 to 20 excellent examples per document type, strip personal and confidential details, add short style notes and store them where Copilot, Gemini, Nextcloud Assistant or an n8n workflow can find them.

Problem & value

  • Today: good documents sit in individual mailboxes; new staff copy "some old offer" that may be outdated or contain another customer's data.
  • AI value: with a curated reference set, drafts reproduce structure, tone, standard clauses and level of detail. Fewer review rounds, faster onboarding and consistent quality.
  • Honest limit: no sourced benchmark exists for time saved; measure it yourself (see KPIs). Office AI products can ground drafts in your own files [1][2].

Variants

  • Sales: offers, proposals, tender answers, follow-up mails.
  • Service and trades: service reports, inspection protocols, handover documents, work orders.
  • Customer service: reply templates for the 20 most frequent request types.

How it works

  1. nominate document types
  2. collect candidates
  3. human selects best (quality criteria)
  4. anonymise & generalise (placeholders)
  5. add metadata + style notes
  6. store in controlled library (SharePoint / Drive / Nextcloud / Paperless-ngx)
  7. used by: Copilot/Gemini reference file | custom assistant | n8n few-shot prompt | RAG index
  8. human reviews every generated draft
  9. feedback updates the library
  • Selection is human work: the AI cannot know which offer won or which letter caused a complaint.
  • Use: two to five matching examples go into the prompt (few-shot); larger libraries use an index (PREP-013, PREP-010).
  • Human-in-the-loop: a named owner approves each item and retires outdated ones; staff always check the generated output.

Data preparation

This use case is data preparation, and it is mostly manual and editorial.

Where examples live: sent mails, personal drives, DMS or file share, CRM/ERP, DATEV folders, paper binders.

Typical gaps: near-identical versions, outdated prices and terms, customer personal data, signatures, confidential conditions, scans without text, no marker of what was "good".

Concrete steps (rough effort for one document type): 1. Choose 3 to 5 document types with highest volume and pain (1 hour workshop). 2. Collect 15 to 30 candidates per type from the staff who write them (2 to 4 hours). 3. Let two experienced colleagues rate and pick the best 5 to 10. Criteria: correct and current, complete, clear, won or praised (2 hours). 4. Anonymise: replace names, addresses, prices and project details with placeholders such as {{Kunde}}, {{Preis}}. See PREP-022; about 20 to 40 minutes per document. 5. Write a half-page style note per type: tone, Sie/Du, length, mandatory clauses (terms: PREP-015). 6. Tag metadata (type, product, language, date, owner, valid until) per PREP-010. 7. Digitise scans if needed (PREP-006); Paperless-ngx does OCR and tags. 8. Restrict edit and read rights (PREP-023); do not index private mailboxes. Set a 6-month review date and feedback route (PREP-021, PREP-017).

Tools SMBs already have: Word and SharePoint or OneDrive, Google Drive and Docs, Nextcloud, Excel for the catalogue, Paperless-ngx, DATEV Dokumentenablage, Power Automate or n8n to copy approved items.

Effort: for 5 document types, plan roughly 3 to 6 working days of mixed staff time, spread over 3 to 4 weeks. This is the real cost driver; the tools cost little. See GUIDE-003.

Implementation paths

Which path for whom

  • SMB — Path A: the office suite or a mature assistant product with reference-file or notebook features is already licensed and the process is standard. The library is just folders and metadata.
  • Enterprise — Path B: many departments, many languages and a shared internal AI platform: an API-based assistant with retrieval over the central library and access control.
  • Public — Path C: letters and notices contain citizen data and need sovereign operation, so a self-hosted index and small model on own or German-cloud infrastructure.
  • Microsoft 365 Copilot: Copilot can ground drafts in files the user may access [1]; check the EU Data Boundary terms for your tenant. Requires clean permissions first (PREP-023).
  • Google Workspace with Gemini: Gems and NotebookLM can use uploaded example documents as grounding, subject to your data-region settings [2].
  • Nextcloud Hub with Assistant (EU-hostable, can run with local models) and Mistral Le Chat (EU provider; check current document-library features) are alternatives.
  • Choose A when: under ~500 examples and no ERP integration needed.

Path B — Build on cloud APIs

  • Use an API model with EU region or EU provider (Azure OpenAI in EU regions, Mistral API, Vertex AI EU) plus an n8n or Power Automate flow: request type → fetch matching examples and style note → draft → human review in Word or mail client.
  • Choose B when: drafts must be created from ERP or CRM data (offer lines, customer history).

Path C — Self-host (open source)

  • Library in Nextcloud or Paperless-ngx, self-hosted n8n, small open-weight model via Ollama (7 to 14B) on HW1/HW2. 2 to 4 days setup plus patching. Good for sensitive content (legal, health, trade secrets).
  • Small models suit template-based drafting; test long reports with your own material.

Path D — Custom training

Almost never; few-shot examples and retrieval cover the need.

State of the art (2026)

  • Few-shot prompting with 2 to 5 good examples remains the cheapest way to match style.
  • Curation quality matters more than model choice.
  • Failure modes: the model copies outdated prices or clauses from an example; leaks details of another customer if examples were not anonymised; mixes tone of different examples; hallucinates facts not in the input. Placeholders and a human check mitigate this.

Shared resources & combinations

Effort, cost & KPIs

  • Path A: 3 to 6 days staff time; add-on licences are priced per user and month (check current lists).
  • Path B: plus 3 to 10 days build.
  • Path C: plus 2 to 4 days setup; HW1 suffices, HW2 is faster.
  • KPIs: time to first acceptable draft, share of drafts accepted with minor edits, review rounds per document, onboarding time for new staff, number of outdated items found per review.

People & change

Involve early: department leads (own quality), the best writers as selectors, IT for library and access, the DPO for anonymisation rules, data owners of CRM and mail data. Management sets scope and time budget.

Works council: depends. Collecting documents is not monitoring, but picking "best" examples can implicitly rate authors and Copilot-type tools log usage, so §87(1) no. 6 BetrVG is likely for the tool rollout; inform the works council in time (§90(1) no. 3). Agree that authors are not named and ratings are not used to assess staff (GUIDE-004). Public bodies involve the Personalrat.

Data protection: offers and letters contain personal data (GDPR). Anonymise before indexing, sign an AVV (Art. 28 GDPR) with any provider, and involve the DPO. A DPIA is normally not needed for a curated, anonymised library, unless mailboxes are indexed at scale (GUIDE-005).

AI literacy: under Art. 4 AI Act, providers and deployers must take measures that support the development of AI literacy of staff (no guaranteed level). A one-hour session on checking AI drafts fits well (GUIDE-006).

Pilot: one team, two document types, 4 weeks, one AI champion (GUIDE-008, GUIDE-009). Concerns such as "my work is copied" or "AI replaces me": authors are not named, selecting and reviewing stay human.

Risks, regulation & compliance

  • EU AI Act: minimal risk for drafting support; Art. 4 AI literacy applies. Art. 50 transparency duties (apply since 2 Aug 2026) can be relevant if AI-generated text is published to inform the public on matters of public interest, unless it has human review and editorial responsibility; plan human review (GUIDE-014).
  • GDPR: personal data in examples; anonymise, minimise, restrict access, define retention. Customer contract terms may be confidential: check NDAs before reuse.
  • Quality risk: outdated examples spread errors; assign an owner and review dates.

Getting started

  1. Pick two document types and an owner each; run a 1-hour workshop (GUIDE-001).
  2. Collect, rate and choose the best 5 to 10 examples per type.
  3. Anonymise, add placeholders and write the style note.
  4. Store in a controlled folder with metadata and test with Copilot, Gemini or an n8n flow on 10 real requests.
  5. Review results with users, retire weak examples and set a 6-month review date.

Sources

  1. Microsoft Learn — Microsoft 365 Copilot overview — https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-overview
  2. Google Workspace — Gemini and AI features — https://workspace.google.com/solutions/ai/
  3. Paperless-ngx documentation — https://docs.paperless-ngx.com/
  4. n8n documentation — https://docs.n8n.io/
  5. Nextcloud — Assistant — https://nextcloud.com/assistant/
  6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj