Report, proposal & offer writing assistant

Writing reports, proposals and offers takes up a large share of knowledge workers' time. It is repetitive: the same structure, the same boilerplate and the same customer or project data, typed again and again. A writing assistant combines a general LLM with the company's templates, past documents and live CRM/ERP/project data. It produces a well-structured first draft in minutes, and the human adds judgement, checks facts and signs off. Prices, quantities and legal terms are never invented by the model. They come from master data and approved text blocks. The typical gain is faster turnaround and more consistent documents, especially for less experienced writers.

Problem & value

• Today: consultants, sales engineers, project managers, tradespeople and officials start from an old Word file. They copy and paste blocks, re-key customer and project data and polish wording late in the evening. Offers go out late, quality varies between authors, and good past text is hard to find.
• AI value: first drafts from notes or voice memos, automatic reuse of the best past sections (retrieval), consistent structure and tone, and quick variants (short/long, technical/executive, German/English). Staff move from rough-drafting to editing and judgement.
• Evidence: in a preregistered experiment with 444 college-educated professionals doing occupation-specific writing tasks (reports, memos, analyses), ChatGPT access cut time taken by 0.8 standard deviations and raised quality by 0.4 SDs. Low performers benefited most, and work shifted from rough-drafting to idea generation and editing [1]. Vendor-reported figures for response tools exist (e.g. Loopio claims around 50% time savings for RFP teams [7]), but they are self-reported and refer to the neighbouring RFP use case (LANG-006).

Variants

• B2B sales proposals (Mittelstand, IT, services): proposal with needs analysis, solution, scope, project plan and references, generated from the CRM opportunity and discovery notes.
• Craft and trades offers (Handwerk, construction): an offer (Angebot) or cost estimate (Kostenvoranschlag) built from site notes, photos, measurements and voice memos. Line items come from the article/service master data. Example: Plancraft turns notes, voice recordings, images and price lists into an editable offer draft [6].
• Consulting and professional services: engagement proposals, statements of work and final/interim reports reusing a library of methods, CVs and case references.
• Periodic business reports: monthly/quarterly management reports, project status reports and board memos combining structured KPIs with narrative. When the narrative is generated mostly from data, see DATA-002.
• Inspection, survey and expert reports: site, audit, maintenance and appraisal reports drafted from field notes and checklists, often combined with dictation (SPEECH-005).
• Public sector: project reports, funding progress reports and internal decision papers. Legislative drafting → LANG-034. Grant and tender responses → LANG-006.
• Private use: freelancers' quotes, reports for associations and clubs, and personal project write-ups. Students' academic work has its own integrity rules.
• Out of scope (own files): RFP/tender responses (LANG-006), make-to-order pricing from CAD geometry (AGENT-027), product configuration and CPQ recommendations (REC-010), clinical/regulatory submission documents (HLTH-017), credit memos (AGENT-028), technical documentation (LANG-022), FMEA/8D (SCI-025), slide decks (GEN-005).

How it works

1. trigger (new CRM opportunity / project period end / site visit)
2. gather context (CRM, ERP price & article data, PM tool, notes, transcripts, photos)
3. retrieve similar past documents & approved text blocks (hybrid search + reranker)
4. outline generation against the document template (sections, mandatory parts)
5. section-by-section drafting with citations to sources
6. deterministic insertion of prices, quantities, dates, legal terms (from ERP/CPQ/templates)
7. consistency & completeness checks (open points, unsupported claims, totals)
8. human review & editing in Word/Docs
9. approval (four-eyes for offers above threshold)
10. export to corporate template (DOCX/PDF), filing in CRM/DMS

• LLM drafting: a general LLM writes the narrative. A reasoning model helps with long, structured documents, such as building a project plan from requirements or checking an argument.
• Retrieval layer: an embedding index plus reranker over won proposals, reference projects, CVs and text-block libraries. This is what makes drafts specific to the company instead of generic.
• Tool calls / agents: connectors (MCP, REST) read the CRM opportunity, ERP article prices and project KPIs. Numbers enter the document through tools or template fields, not free generation.
• Parsing: old proposals and price lists in PDF/scan are ingested with document parsing (DOC-014).
• Human-in-the-loop: the author edits every draft. Offers are legally relevant (binding under §145 BGB once sent). Price, discount and terms approval stays with authorised staff. Reports to boards or regulators are signed off by the responsible person.

Data preparation

Needed: article, service and price master data (ERP, CPQ, price lists in Excel), customer and opportunity data (CRM), templates and boilerplate text blocks, past reports, offers and proposals (DMS, file shares, mail attachments, partly scanned or paper), and project data from PM tools.

Typical gaps: price lists in messy spreadsheets, several versions of the same offer without a "won/lost" label, boilerplate scattered across personal files, expert knowledge only in people's heads, and access rights on project folders that are too open or unclear.

Preparation steps: - Turn price and article data into tidy tables with one source of truth (see PREP-011, PREP-005). - Select 20 to 50 good past documents as reference library, label them (won, lost, topic) and OCR scans (see PREP-016, PREP-020, PREP-006, DOC-014). - Consolidate templates and text blocks into one approved set (see PREP-013, PREP-014). - Restructure folders and tags so retrieval finds the right version (see PREP-009, PREP-010); clean up permissions (see PREP-023).

Rough share: 50 to 70 percent of the effort. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path B. A good offer or report needs your own price lists, CRM data and templates; generic tools do not know them, so an API-based workflow (for example with n8n) fits better than a boxed product.
• Enterprise: Path C. Offers and proposals contain pricing and strategy, which are trade secrets, and volume makes seat costs high; a shared internal AI platform with RAG fits.
• Developer: Path B. Cloud APIs with document and CRM connectors let you build the pipeline quickly.
• Private: Path A. For personal reports and applications a ready-made assistant is enough.
• Public sector: Path C. Reports and tenders contain sensitive data, and sovereign EU or German cloud or self-hosting is required by procurement and sovereignty rules.

Path A — Buy (off-the-shelf)

• Office copilots: Microsoft 365 Copilot in Word drafts from prompts, outlines, notes or referenced files, reuses templates and existing documents, and can add citations. It offers model choice (including Claude models) [3]. M365 data is processed in the EU Data Boundary for EU-tenant customers [9]. Google Gemini in Workspace is the equivalent for Google shops.
• EU chat/agent platforms (marked EU): Langdock (Berlin; EU hosting, own-cloud or on-prem from 1,000+ seats, model-agnostic, ISO 27001/SOC 2) [4]. Mistral Le Chat Enterprise (FR). Both let you build a "proposal assistant" agent with your templates and knowledge files.
• Proposal/document platforms: PandaDoc, Qwilr and Proposify for sales proposals with CRM integration (HubSpot, Salesforce), content libraries and e-signature. Check where data is hosted.
• Trade software with AI offers (DE): Plancraft (offers from voice, photos and price lists) [6] and other craft ERP vendors adding similar features.
• Writing polish: DeepL Write (DE vendor) for style and tone modes [8]. See LANG-011.
• Choose A when: documents live in M365/Google, volume is moderate, and the main need is faster first drafts.

Path B — Build on cloud APIs

• Models: Mistral Medium 3.5 / Large 3 via La Plateforme (EU provider) [5]. Claude, GPT and Gemini via EU regions on AWS Bedrock (Frankfurt), Azure OpenAI/Foundry (EU data zones) and Google Vertex AI (europe-west).
• Pattern: template-driven generation. A JSON outline schema fills a DOCX template (python-docx/docxtpl, Carbone). Retrieval goes over a vector store (pgvector, Qdrant). Tools read CRM/ERP (HubSpot, Salesforce, SAP Business One, Dynamics 365 Business Central) through APIs or MCP servers.
• Orchestration: n8n or Power Automate for the trigger (e.g. "opportunity reached stage proposal"). LangGraph or a vendor agent SDK for multi-step drafting and checking.
• Choose B when: you want offers generated directly from CRM/ERP with your own price logic, numbered text blocks and corporate design.

Path C — Self-host (open source)

• Models: Qwen3-30B-A3B-Instruct-2507 (Apache 2.0, 262k native context) [10]. Mistral Small 4 (Apache 2.0) [5]. Larger options (Mistral Large 3, Apache 2.0 [5]; GLM-family or Qwen3 235B-class) for higher German writing quality.
• Stack: vLLM with JSON-schema structured outputs for outlines and line items [11]. Open WebUI or LibreChat as the front end. Qdrant/pgvector plus a BGE/multilingual-e5 embedding model and a reranker. Docling for ingesting old proposals. n8n self-hosted for triggers.
• Hardware: HW2 (one 24–32 GB GPU) runs a quantised ~30B MoE model for a team of writers. HW3 for 70–120B-class models, longer documents and many concurrent users.
• Ops: moderate. Model updates, template maintenance, index refresh and access control on the reference library (not every user may see every past offer).

Path D — Custom training

Rarely worth it. Retrieval of good examples plus a clear style guide usually gives company-specific tone. Fine-tuning (LoRA on an open model) can pay off for high-volume, highly standardised documents, such as thousands of near-identical inspection reports. It needs several hundred to a few thousand approved final documents, ideally paired with the input notes they were written from.

State of the art (2026)

• Long-context frontier models (Claude, GPT, Gemini, Mistral Large families) write fluent German and English long-form text. The bottleneck is no longer wording but grounding: correct facts, numbers and commitments.
• Agentic drafting in the office suite: Copilot in Word and similar tools now edit in place, pull in referenced files, emails and meetings, and add citations [3]. Platforms expose MCP connectors, so assistants such as Claude or ChatGPT can query response libraries directly [7].
• Multimodal input: offers from voice memos and site photos are shipping in trade software [6].
• Structured generation: schema-constrained decoding (vLLM, xgrammar/guidance backends [11]) makes outlines and line items machine-checkable before rendering.
• Limits and failure modes: hallucinated references, capabilities or certifications. Invented or mis-summed prices. Copying confidential content from another customer's past proposal (retrieval leakage). Bland, generic "AI tone". Overly long text. Outdated boilerplate. Always keep a source trail per section and recompute totals deterministically.

Shared resources & combinations

• Same LLM + RAG stack (llm-general, embedding-text, vector-db, reranker): LANG-001 marketing copy, LANG-002 email drafting, LANG-006 RFP responses (shares the reference library), LANG-008 summarisation, LANG-011 style rewriting, KNOW-002 company RAG assistant. One HW2/HW3 inference server and one index can serve all of them.
• Upstream inputs: SPEECH-001 meeting summaries and SPEECH-005 dictation (notes to report), KNOW-015 deep research (market sections), DATA-002 data narratives (KPI sections), DOC-014 parsing of legacy proposals.
• Downstream: LANG-009 translation of offers, GEN-005 slides from the proposal, DOC-010 contract drafting once the offer is accepted.
• Specialised neighbours: AGENT-027 RFQ-to-quote (pricing from CAD), REC-010 configuration recommendations feeding offer line items, AGENT-028 credit memos, HLTH-017 regulated medical writing.
• Workflow engine (n8n / Power Automate) shared with AGENT-002 back-office automation.

Effort, cost & KPIs

• Path A: days. Per-user licences for office copilots or chat platforms. Main effort is cleaning templates and building a text-block library.
• Path B: 4–10 weeks for a CRM/ERP-integrated offer generator with templates. API cost per document is usually small compared to staff time.
• Path C: 1–3 months including hardware, index and front end, plus ongoing operations.
• Cost drivers: integration with ERP prices and CRM, template quality, review time, and access control on the reference library.
• KPIs: time from request to sent offer/proposal, author hours per document, share of the draft kept after editing, error rate found in review (prices, names, commitments), win rate and offer conversion, number of offers per sales FTE, report punctuality.

People & change

Involve management (which documents may be AI-drafted), the sales or project lead as process owner, the data owners of ERP, CRM and the DMS, IT for the interfaces and the DPO for customer and employee data in reference documents. Legal or compliance should confirm that terms and conditions and legal clauses come only from approved blocks.

Pilot group: two to four experienced offer writers, who define the checklist and decide which blocks are "locked". Keep prices, quantities and legal terms always from master data; the human signs off every document.

Works council: depends. A pure drafting assistant on one workstation may not trigger §87(1) no. 6 BetrVG, but platforms that log individual use (for example M365 Copilot) do, according to the BAG's objective-suitability test. §90(1) no. 3 requires early information in any case. See GUIDE-004. Public sector: staff council.

Data protection: AVV (Art. 28 GDPR) with providers; a DPIA only if sensitive personal data is processed (GUIDE-005). AI literacy: support staff with training on fact-checking and prompts (Art. 4 AI Act, GUIDE-006). Typical concern: "the AI makes the offer wrong." Address it with visible source references and an open-points checklist; run the pilot per GUIDE-009.

Risks, regulation & compliance

• EU AI Act: minimal risk in the typical form. Drafting business documents is not an Annex III category, and the AI literacy duty (Art. 4) applies. Art. 50(4) disclosure applies only to AI-generated text published to inform the public on matters of public interest, and not when a person has reviewed it and holds editorial responsibility. Art. 50(2) machine-readable marking is a provider (vendor) duty, with exceptions for systems that mainly assist standard editing. Art. 50 applies from 2 Aug 2026 [2][12]. Digital omnibus, reg. 2026/1744: systems already on the market before 2 aug 2026 have until 2 dec 2026 for art. 50(2) marking.
• GDPR: CRM data, contact persons, CVs of team members and customer project details are personal data. You need a DPA with the vendor, EU hosting, no training on customer data, role-based access to the reference library, and deletion/retention rules.
• Contract & commercial law (DE): an offer is binding once sent (§145 BGB), and a cost estimate has legal consequences if exceeded significantly (§649 BGB). Prices, terms and delivery dates must come from controlled sources and be approved by authorised staff. Accepted offers and related correspondence are commercial letters with retention duties (§257 HGB, GoBD).
• Confidentiality & IP: no reuse of one customer's confidential content in another's proposal. Check NDAs before indexing. Mark AI use where clients or funders require it.
• Mitigations: source citations per section, deterministic price/total calculation, mandatory review checklist, a blocklist of unverified claims (certifications, references), and an audit log of draft vs. final.

Getting started

1. Pick one high-volume document type (e.g. standard offer or monthly project report) and measure current hours per document and turnaround time.
2. Clean up 1–3 master templates and build a library of approved text blocks and 20–50 good past documents (anonymised where needed).
3. Pilot with the office copilot or an EU platform you already have (Path A). Create a dedicated assistant with template, style guide and library.
4. Connect price and article data from ERP/CPQ so numbers are inserted, not generated. Add a four-eyes approval step for offers.
5. After 4–6 weeks compare hours, error rate in review and win rate. Then decide on a deeper CRM/ERP-integrated build (Path B/C).

Sources

1. Noy, S. & Zhang, W. — Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence (MIT working paper; later published in Science 2023) — https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf
2. EU AI Act, Article 50 — Transparency obligations (artificialintelligenceact.eu) — https://artificialintelligenceact.eu/article/50/
3. Microsoft Support — Draft and add content with Copilot in Word — https://support.microsoft.com/en-us/office/draft-and-add-content-with-copilot-in-word-069c91f0-9e42-4c9a-bbce-fddf5d581541
4. Langdock — platform overview (EU hosting, deployment options) — https://langdock.com/
5. Mistral AI — Models overview (versions & licences) — https://docs.mistral.ai/getting-started/models/models_overview/
6. Plancraft — craft business software with AI offer creation — https://plancraft.com/
7. Loopio — Response Intelligence and MCP connector (vendor-reported figures) — https://loopio.com/
8. DeepL Write — https://www.deepl.com/en/write
9. Microsoft — What is the EU Data Boundary? — https://learn.microsoft.com/en-us/privacy/eudb/eu-data-boundary-learn
10. Qwen3-30B-A3B-Instruct-2507 model card — https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507
11. vLLM — Structured outputs — https://docs.vllm.ai/en/latest/features/structured_outputs.html
12. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj