Email & business correspondence drafting

Knowledge workers spend a large part of their day reading and writing email, and Mittelstand companies still send many formal letters (offers follow-ups, complaints, reminders, replies to authorities). An LLM writing assistant turns a few bullet points or the incoming thread into a ready-to-edit draft in the right tone and language, pulls facts from CRM or ERP, and applies the house style. The human stays the author: they review, edit and press send. In 2026 this is a commodity feature of every major mail client, so the work lies in grounding, house style, data protection and adoption.

Problem & value

• Today: staff write replies from scratch or copy text modules, look up order or contract data in other systems, and spend time on tone ("firm but friendly"), foreign-language mail and formal letters. Quality varies widely between employees.
• AI value: faster first drafts, consistent tone and terminology, easier writing in foreign languages, less effort for people with dyslexia or dyspraxia, and fewer missing details when the draft is grounded in business data.
• Measured evidence (modest but real):
• A Microsoft Research field experiment with more than 6,000 workers at 56 firms found that Microsoft 365 Copilot users spent about 30 minutes less per week reading email [1].
• In the UK cross-government Copilot experiment (20,000 staff, 12 organisations, Oct–Dec 2024), users self-reported an average of 26 minutes saved per day. Drafting documents saved the most time (24 minutes); communication tasks saved less but saved it often [2].
• In a Stanford Health Care pilot of AI-drafted replies to patient messages (162 clinicians), drafts were used for about 20 % of replies. Task load and work exhaustion scores fell significantly, but reply time did not change measurably [3]. Lesson: drafting assistants reduce mental load more reliably than they reduce stopwatch time.

Variants

• Individual knowledge worker (enterprise): Copilot in Outlook, Gemini in Gmail; "help me write / help me reply", tone and length controls, coaching [4][5].
• Customer service & sales: grounded replies to customer emails with order, delivery or contract data. Templates are filled by the LLM, and an agent approves the draft. Fully automated sending belongs to AGENT-001. Marketplace and shop replies belong to AGENT-020.
• Online review responses (SMB): public replies to Google, OTA or marketplace reviews are the same drafting task, with brand voice and escalation rules. The specialised use case is LANG-038 (review sentiment: LANG-013).
• Formal business letters (Geschäftsbriefe): DIN 5008 layout. Mandatory company details (e.g. §35a GmbHG requires legal form, seat, register court and number, and managing directors on business letters of any form) are added from a template, never by the LLM [11].
• Multilingual correspondence: draft in German, send in English, French or Polish. Optionally, run a terminology-aware translation step (LANG-009).
• Public sector: replies to citizen enquiries and standard letters. Plain-language rewriting (LANG-010). Binding administrative decisions remain a separate, human-signed procedure.
• Healthcare: patient-message replies inside the EHR have their own safety review → HLTH-030.
• Private use: complaints, cancellations, letters to landlords or authorities → PERS-008; cover letters → PERS-012. Privacy-focused options include Proton Scribe, which can run locally in English [9].

How it works

1. trigger (user opens thread / writes notes / ticket arrives)
2. context assembly: thread + attachments, sender record from CRM/ERP, relevant policy or text modules (retrieval over templates & past approved replies)
3. PII / secret filter and prompt-injection guard on inbound content
4. LLM draft with system prompt = house style, salutation rules, language, length, signature
5. optional translation / plain-language pass
6. checks: facts vs. source data (prices, dates, order numbers), forbidden promises, mandatory legal footer from template
7. draft placed in the mail client / ticket UI ← HUMAN reviews, edits, sends
8. log: draft vs. sent text (quality feedback, style examples)

• Core model: a general instruction-tuned LLM. Mid-size models are enough for most mail. Frontier models help with long threads, negotiation tone and legal nuance.
• Grounding: retrieval over approved text modules and past replies (embeddings + vector index), plus read-only tool calls to CRM/ERP. This is what turns generic text into a correct reply.
• Style: a style guide in the system prompt plus a few examples is usually enough. Fine-tuning is rarely needed.
• Human-in-the-loop: the human sends by default. Auto-send only for narrow, low-risk categories, and then it is AGENT-001, not this use case.

Data preparation

The data is the mail traffic itself (mailboxes, shared inboxes), templates and text modules (Textbausteine) in Word or on file shares, a style guide if one exists, and context from CRM, ERP or the ticket system (order status, contract data). Many firms also keep letters as paper copies or scans.

Typical gaps: no written house style, text modules scattered across personal folders, outdated templates, customer data in several systems, and mailbox permissions that are too broad, so an assistant might surface mails the user should not see.

Preparation steps: - Collect and curate a small set of current templates and 10 to 20 good sent mails as tone examples (see PREP-016, PREP-013). - Write a short style guide and a list of company terms (see PREP-015). - Clean up mailbox and file permissions before rolling out an assistant with access (see PREP-023); define what may go into which tool (see PREP-024). - Connect CRM or ERP via export or API only where drafts need facts (see PREP-005).

Rough share: about 15 to 25 percent of the effort; the pure writing assistant works almost out of the box. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path A. Copilot, Gemini or an EU-hosted mail assistant is a mature standard product at SMB prices, and e-mail is a standard process.
• Enterprise: Path C. Mail contains customer and contract data at scale, so a shared internal AI platform with open-weight models avoids seat costs and keeps content in-house.
• Developer: Path C. Open-weight models plus an IMAP or add-in integration give full control over data and prompts.
• Private: Path C. Local apps such as LM Studio or Ollama keep personal correspondence on your own device.
• Public sector: Path C. Sovereignty and procurement rules favour self-hosted or sovereign German cloud models over US-based services.

Path A — Buy (recommended for most SMBs)

• Built into the mail suite: Microsoft 365 Copilot in Outlook (draft, tone/length, coaching; Microsoft 365 is covered by the EU Data Boundary for EU/EFTA tenants) [4][6]; Gemini in Gmail (Help me write, polish, contextual smart replies) [5].
• EU-based AI workspaces: Langdock (Berlin; EU hosting, Outlook add-in, on-prem from 1,000 seats, ISO 27001/SOC 2) [7]; Mistral's Le Chat/Vibe (France). Other German "company GPT" platforms follow the same pattern.
• Writing/rewriting tools: DeepL Write (Germany; style rules, Outlook add-in and browser extensions) [8]. Grammarly is the US-based alternative.
• Service desks: AI reply drafting in Zendesk, Freshdesk, Salesforce Service Cloud and HubSpot, grounded in the ticket.
• Choose A when: the organisation already runs M365 or Workspace, or wants an EU-hosted chat platform with an Outlook add-in. The budget question is per-seat licences against actual usage.

Path B — Build on cloud APIs

• Models: Mistral Medium 3.5 / Mistral Small 4 via La Plateforme (EU provider) [10]; Claude, GPT or Gemini via EU-region endpoints (AWS Bedrock Frankfurt, Azure EU, Vertex AI europe-west).
• Integration: Outlook add-in (Office.js) or Microsoft Graph, Gmail API, IMAP for other mail systems; ticket-system webhooks; n8n or Power Automate for "draft on arrival".
• Grounding: read-only CRM/ERP tools exposed via MCP or function calling; approved text modules in a vector index (pgvector, Qdrant).
• Choose B when: replies must use internal systems (order status, contract terms), or you need one house style across many channels.

Path C — Self-host (open source)

• Models: Mistral Small 4 (Apache 2.0), Ministral 3 8B/14B, Gemma 4, Qwen3.5/Qwen3.6 (strong multilingual), gpt-oss-20b/120b [10][12]. For German business mail, test Mistral and Gemma models first. Small models (≤ 8B) work for autocomplete and polishing.
• Stack: vLLM or Ollama for serving; Open WebUI or LibreChat as the user front end; a self-built Outlook add-in or a Thunderbird extension calling the internal endpoint; Presidio or similar for PII handling; n8n for mailbox automation.
• Hardware: HW2 (one 24–32 GB GPU) serves a 14–30B quantised model for dozens of concurrent users. HW3 for 100B-class models (gpt-oss-120b, Mistral Medium 3.5) or hundreds of users.
• Choose C when: correspondence contains confidential or professional-secrecy data (lawyers, tax advisers, physicians, §203 StGB), or policy forbids external processors.

Path D — Custom training

Rarely worth it. A LoRA fine-tune of a small open model on a few thousand approved sent replies can lock in a very specific house style or domain phrasing, or let a smaller, cheaper model replace a larger one at high volume (e.g. insurers, utilities, public agencies with standard letters). Check first that retrieval plus few-shot examples are really insufficient. Training on sent mail requires a GDPR legal basis and pseudonymisation (DOC-018).

State of the art (2026)

• Commodity in clients: drafting, rewriting, tone controls and contextual reply suggestions ship with Outlook, Gmail and Proton Mail [4][5][9]. Differentiation now comes from grounding in company data and from agentic extensions (look up an order, propose a meeting slot, attach the right document).
• Open models are good enough: current open-weight families (Mistral Small 4, Gemma 4, Qwen3.5, gpt-oss) write fluent German, French and English business mail on a single GPU, which makes on-prem drafting realistic for the Mittelstand [10][12].
• Evidence is mixed: controlled studies show modest time savings and reduced burden rather than dramatic productivity gains [1][3]. Self-reported savings are larger [2].
• Limits & failure modes: invented facts (prices, dates, commitments) when not grounded; over-long, generic "AI tone" that recipients notice; wrong formality (du/Sie) or cultural register; prompt injection via inbound emails that instruct the assistant; accidental disclosure of data from other threads in retrieval; automation bias (users send unread drafts).

Shared resources & combinations

• Same LLM endpoint (llm-general, HW2/HW3): LANG-011 proofreading & brand voice, LANG-009 translation, LANG-010 plain language, LANG-003 reports & offers, LANG-038 review replies. One self-hosted model server plus one style guide serves all of them.
• Same retrieval stack (embedding-text + vector-db): KNOW-002 RAG assistant. Approved text modules and FAQs are a shared knowledge base.
• Upstream/downstream: LANG-012 classifies and routes incoming mail before drafting. AGENT-001 adds triage and auto-send. CONV-015 personal assistant combines drafting with calendar and tasks. AGENT-020 for shop replies.
• Shared guardrails: SEC-010 prompt-injection defence and SEC-017 DLP apply to every mail-reading assistant; DOC-018 redaction for training data.

Effort, cost & KPIs

• Path A: days. Cost is per-seat licences (suite add-on or AI workspace). Main effort: style guide, training users, deciding who gets a licence.
• Path B: 3–8 weeks for an Outlook add-in or service-desk integration with CRM grounding. Token cost per email is small compared with staff time.
• Path C: 1–2 months, one GPU server plus ongoing model updates and monitoring.
• KPIs: share of replies that start from an AI draft; edit distance between draft and sent text; handling time per email or ticket (measured, not only surveyed); first-contact resolution; customer satisfaction for replied tickets; factual error rate found in sampled audits; licence utilisation.

People & change

Involve management (AI policy, budget), the department leads, IT and IT security (mailbox access, tenant settings) and the DPO early, because mail holds personal data of customers and staff. Pick a pilot group of 5 to 10 heavy mail users from different teams, run it for four to six weeks and collect examples of good and bad drafts.

Works council: yes in practice. Mail assistants such as M365 Copilot log user activity, and the BAG holds that objective suitability for monitoring is enough for co-determination under §87(1) no. 6 BetrVG. §90(1) no. 3 also requires information in good time, and §80(3) lets the council engage an AI expert. Public sector: staff council (Personalrat). A works agreement on AI use is the usual solution (GUIDE-004).

Data protection: a DPIA is usually not required for pure drafting, but check it when mail content is analysed systematically; an AVV (Art. 28 GDPR) is mandatory (GUIDE-005). AI literacy: support staff with short training on confidential data, checking facts and tone (Art. 4 AI Act, GUIDE-006). Typical concerns are surveillance and de-skilling: state clearly that no performance evaluation happens, and that the author always reviews and sends.

Risks, regulation & compliance

• EU AI Act: minimal risk. Drafting correspondence is not an Annex III use, and AI literacy (Art. 4) applies to staff. Art. 50(2) marking duties fall on providers of text-generating systems from 2 Aug 2026 (Digital Omnibus, Reg. 2026/1744: systems already on the market before 2 Aug 2026 have until 2 Dec 2026 for Art. 50(2) marking), with an exception for assistive editing [13][14]. Art. 50(4) disclosure for deployers covers text published to inform the public on matters of public interest, not private correspondence, and does not apply after human editorial review. Customer-facing auto-reply bots trigger the Art. 50(1) chatbot disclosure → AGENT-001/CONV-001.
• GDPR: emails almost always contain personal data. You need a data processing agreement, EU processing, a no-training-on-customer-data clause, retention settings and a DPIA where special categories occur (health, HR). Respect purpose limitation before using sent mail as training data.
• Professional secrecy: lawyers, tax advisers and physicians (§203 StGB) need contractual confidentiality obligations with the provider, or Path C.
• Commercial law: mandatory business-letter details (§35a GmbHG, §37a HGB) and signatures come from templates [11]. Letters that create legal effects (offers, terminations, dunning) need human approval. Archive commercially relevant mail per GoBD/HGB retention.
• Security: treat inbound mail as untrusted input to the LLM. Disable autonomous tool actions triggered by email content, or require confirmation for them.
• Works council: in Germany, rolling out Copilot-type tools that log usage typically needs a works agreement (§87(1) no. 6 BetrVG).

Getting started

1. Pick 2–3 high-volume correspondence types (e.g. order-status questions, complaint replies, offer follow-ups) and measure today's handling time and quality.
2. Write a one-page style guide (salutation, du/Sie, length, sign-off, forbidden promises) and collect 10–20 good example replies.
3. Pilot the suite assistant you already license (Copilot/Gemini) or an EU workspace such as Langdock with 10–20 users for four weeks. Compare handling time and draft edit distance.
4. If replies need business data, build a grounded Path B/C add-in with read-only CRM/ERP lookups and a prompt-injection guard.
5. Agree on rules for human review, data protection (DPA, DPIA) and works council before scaling.

Sources

1. Dillon, Jaffe, Peng, Cambon — "Early Impacts of M365 Copilot" (Microsoft Research field experiment), arXiv:2504.11443 — https://arxiv.org/abs/2504.11443
2. UK Government — Microsoft 365 Copilot Experiment: Cross-Government Findings Report (June 2025) — https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report
3. Garcia P. et al. — "Artificial Intelligence–Generated Draft Replies to Patient Inbox Messages", JAMA Netw Open 2024;7(3):e243201 — https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2816494
4. Microsoft Support — Draft an email message with Copilot in Outlook — https://support.microsoft.com/en-us/office/draft-an-email-message-with-copilot-in-outlook-3eb1d053-89b8-491c-8a6e-746015238d9b
5. Google Workspace — Gemini in Gmail — https://workspace.google.com/products/gmail/ai/
6. Microsoft — What is the EU Data Boundary? — https://learn.microsoft.com/en-us/privacy/eudb/eu-data-boundary-learn
7. Langdock — product overview (EU hosting, Outlook add-in, on-prem) — https://www.langdock.com/
8. DeepL Write — https://www.deepl.com/en/write
9. Proton — Proton Scribe writing assistant — https://proton.me/support/proton-scribe-writing-assistant
10. Mistral AI — Models overview & deprecations — https://docs.mistral.ai/getting-started/models/models_overview/
11. § 35a GmbHG — Angaben auf Geschäftsbriefen — https://www.gesetze-im-internet.de/gmbhg/__35a.html
12. Ollama model library (availability of open-weight models) — https://ollama.com/library
13. EU AI Act Article 50 (transparency obligations) — https://artificialintelligenceact.eu/article/50/
14. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj