AI-personalised mail merge & serial letters

Classic mail merge puts a name and an address into a fixed text. Many real letters need more: a reminder that mentions the open item, a notice that fits the contract type, an information letter whose second paragraph differs by customer group. An AI step writes only those passages from the row data, a person reviews the drafts, and the existing merge or mail tool does the rest. It is a days-long project with tools an SMB already owns, and it is mostly a data-preparation exercise.

Problem & value

• Today: offices send serial letters and mails (price changes, contract notices, membership reminders, event invitations, supplier requests) by merge fields only, or they write 20 variants by hand. Special cases are either ignored or typed individually, and the recipient list is often an old Excel sheet with mixed spelling and duplicates.
• Why AI helps: the model turns context columns into one or two fitting sentences ("Your service contract for the heating system ends on 31 March; we offer a visit in February"), keeps a consistent tone and can write in the recipient's language. The fixed legal text stays untouched.
• Benefit: fewer hand-written variants and more relevant letters. No reliable cross-industry figure exists, so measure on your own mailing: minutes per letter before and after, share of drafts edited, replies and returned mail.
• Not in scope: marketing campaign personalisation, segmentation and send-time optimisation (REC-006). Here it is the office tool for notices, reminders and information letters.

Variants

• Letters: Word template to DOCX or PDF, printed or sent through a letter-dispatch service.
• Emails: one individual email per recipient from a shared mailbox, with approval before sending.
• Association / club / school: member and parent letters; special care for minors' data.
• Tradespeople and property managers: maintenance notices, appointment proposals, tenant information.
• Public bodies: citizen information letters with a fixed legal core and AI only for explanatory sentences.
• Private use: club newsletters or invitations with a local model (LM Studio, Ollama).

How it works

1. recipient list (Excel / Sheets / CRM export)
2. data check (duplicates, empty fields, address format, opt-outs, excluded persons)
3. per row: AI step writes only the marked passages from the context columns
4. rule check (no amounts/dates not in the row, forbidden phrases, length)
5. merge into template (Word, Google Docs, email draft)
6. human review of all or a sample, status "check" for flagged rows
7. approved send / print; log

• Key AI component: a general or small LLM with a strict prompt: use only the given fields, 1-3 sentences, fixed tone, return "no personal passage" when context is missing. Fixed text, numbers, dates and amounts come from the table, never from the model.
• Human in the loop: the whole batch or a sample (first 20 plus every flagged row) is read before sending. Nothing is sent automatically.

Data preparation

Which data: the recipient table with one row per person or company, clean address and salutation fields, the context columns the AI may use, and a consent/opt-out or exclusion flag.

Where it lives: Excel or Google Sheets, a CRM, DATEV address data, the association's member tool, or exports from a shop or ERP; context remarks are often in free-text columns.

Typical gaps: duplicates ("Müller GmbH" and "Mueller GmbH"), missing or wrong salutation, merged cells, mixed date formats, outdated addresses, remarks with sensitive content (health, disputes) that must not reach the model, and no marker for people who objected to contact.

Preparation steps: 1. Turn the sheet into a proper table with one header row and one row per recipient (1-3 hours; see PREP-011 and AUTO-006). 2. Remove duplicates and normalise salutation, names and dates (Power Query or the Excel "Remove duplicates"; 1-2 hours). 3. Decide per column whether it may be sent to the AI: send only the context needed; drop or pseudonymise free-text remarks (PREP-022, PREP-024). 4. Add an exclusion column (objection, dispute, insolvency, deceased) and filter before the AI step. 5. Write 3-5 gold-standard letters by hand as examples and tone reference (PREP-016) and keep 20 rows as a test set (PREP-018).

Tools: Excel and Power Query, Google Sheets, Word, OpenRefine for messy names, Nextcloud Tables.

Rough share of effort: about half of the project; the AI step is quick once the table is clean. See GUIDE-003.

Implementation paths

Which path for whom

• SMB - Path B: the lists and templates are your own and no product covers every letter type, so a small flow with an EU-hosted model API fits. Switch to Path C if the context columns contain health, HR or other sensitive data.
• Enterprise - Path B: a central workflow platform with approved EU model endpoints, reused by departments; integration with the CRM or ERP.
• Developer - Path B: a short script or n8n flow with structured output against an EU API.
• Public - Path C: citizen data and procurement rules favour local or sovereign hosting; a small local model is enough for short passages.

Path A - Buy (off-the-shelf)

• Microsoft 365: Word mail merge with Excel or an Outlook contact list as data source is built in and well documented [1]. Microsoft 365 Copilot can draft or rewrite text in Word, but it does not fill passages per row inside the merge; check current licensing and the data boundary for your tenant.
• Google Workspace: Docs/Sheets add-ons for mail merge plus Gemini for drafting the template wording.
• Dedicated tools: letter-dispatch services (print and mail from the cloud, several with German data centres) and CRM or association software with serial-letter modules. AI personalisation inside them is usually limited to a text box.
• Choose A when: the letters only need merge fields and a handful of text variants selected by rules. This often suffices and needs no AI.

Path B - Build on cloud APIs (recommended for most SMBs)

• Orchestration: n8n (cloud with EU region or self-hosted), Power Automate, Make. n8n offers a Basic LLM Chain node with system and user messages and an optional output parser [2]; Power Automate can call an HTTP endpoint or AI Builder.
• Flow: trigger by "status = ready" in the sheet, loop over rows, call the model, write the passage back to a new column, then use normal Word merge or an email draft node.
• Models: EU providers such as Mistral (France; check the DPA and subprocessor list on its legal pages [3]), or Azure OpenAI and Google Vertex AI in an EU region. Small models are enough for short passages.
• Effort: 1-3 days for the first flow, about a week of testing with real rows.

Path C - Self-host (open source)

• Stack: n8n Community Edition (source-available, internal business use) or a Python script, Ollama with a small model (for example a Mistral Small or Qwen-class 3-14B), reading and writing a Nextcloud table or CSV, then Word/LibreOffice merge.
• Hardware: HW1 is enough for short passages in a batch of a few hundred rows; run it overnight. A workstation (HW2) speeds up and improves German text quality.
• Ops effort: a few hours per month. Worth it for customer data in regulated sectors, member data of minors or when the same server serves AUTO-008 and AUTO-015 (Baukasten effect).

Path D - Custom training

Not needed. Prompt, rules and 3-5 example letters are enough; fine-tuning on a house style only pays at very large volumes.

State of the art (2026)

• Current models write fluent German business text from a few fields and follow format constraints reliably when given structured output and examples.
• Limits and failure modes: invented facts (dates, amounts, promises) that are not in the row, wrong tone for sensitive cases (arrears, bereavement, complaints), wrong salutation or gender, repeated phrases across hundreds of letters, and leakage of internal remarks into the letter. Mitigate with rule checks, "no personal passage" fallback, excluded categories handled by hand, and a mandatory review step.

Shared resources & combinations

• Same stack (workflow-automation + llm-small): AUTO-008 form to rows, AUTO-007 attachment filing, AUTO-013 contact capture; the clean recipient table comes from AUTO-006.
• Natural combination: AUTO-015 recurring document generation from templates (the merge and PDF step); AUTO-002 for reply drafts when recipients answer.
• Upgrades: campaign personalisation and send-time optimisation in REC-006; redaction of remarks before the AI step in DOC-018.

Effort, cost & KPIs

• Path A: hours for Word merge; Copilot or add-on licence cost per seat.
• Path B: 1-3 days build, about a week pilot; running cost is platform fee plus small per-letter model cost.
• Path C: 3-5 days including server setup, plus light operations.
• KPIs: share of drafts sent unchanged, share flagged "check", minutes per letter, returned or wrongly addressed mail, complaints, replies or payments after reminders.

People & change

Involve early: the department lead (owns the letter and its legal content), end users who write the letters today, IT for accounts and connectors, the DPO, and the data owner of the recipient list (GUIDE-005). Management approves the pilot. Customer service should know that letters are changing.

Works council: depends. Letters to customers and members usually do not trigger co-determination under §87(1) no. 6 BetrVG. It becomes relevant if the flow logs who edited or sent which letter and how fast, or if recipients are employees (staff information letters). The BAG holds that objective suitability to monitor is enough, so agree on "no performance evaluation from flow logs". §90(1) no. 3 requires timely information when work processes are planned with AI (GUIDE-004). Public bodies involve the Personalrat.

Data protection: personal data is processed. Check purpose and legal basis for the letter (contract, legitimate interest, consent for marketing content), sign an AVV (Art. 28 GDPR) with the model and platform provider, involve the DPO, and send only needed fields. A DPIA (Art. 35 GDPR) is normally not needed for simple notices, but check it for special categories or systematic evaluation of people (GUIDE-005).

AI literacy: Art. 4 AI Act requires providers and deployers to take measures that support staff AI literacy. A 30-minute session on reviewing drafts and spotting invented facts is enough (GUIDE-006). Add the use to your AI policy (GUIDE-007).

Humans in control: a person reads and approves before anything is sent; sensitive cases are written by hand. Start with one pilot mailing and a small group (GUIDE-008, GUIDE-009).

Risks, regulation & compliance

• EU AI Act: I classify it as limited in the vocabulary because Art. 50 transparency duties may apply when AI-generated text is published to inform the public on matters of public interest; for individual business letters reviewed by a person the practical duty is usually small (GUIDE-014). Art. 4 applies. Do not reuse the pattern for application rejections or HR selection (Annex III high-risk, obligations from 2 Dec 2027).
• GDPR: data minimisation, purpose limitation, accuracy (the model must not add facts), objection handling and advertising rules (UWG, §7) for marketing content, retention of generated drafts, EU hosting or adequate safeguards.
• Security: API keys in flows, wide sharing of the recipient sheet, prompt injection through free-text columns filled by third parties (GUIDE-013).
• Mitigations: minimal columns to the model, exclusion list, fixed legal text outside the model, review step, logging agreed with the works council.

Getting started

1. Pick one recurring mailing with 50-500 recipients and clean its table (GUIDE-001, GUIDE-003).
2. Write the fixed text yourself and mark the 1-2 passages the AI may personalise; write 3-5 example letters.
3. First try plain merge fields and rules; add the AI step only where it saves real manual variants (GUIDE-002).
4. Build the flow in Power Automate or n8n with a "check" status, inform the DPO and, where applicable, the works council, and sign the AVV (GUIDE-004, GUIDE-005).
5. Run a pilot with 20 rows reviewed in full, measure edits, then decide go or no-go (GUIDE-009).

Sources

1. Microsoft - Use mail merge for bulk email, letters, labels and envelopes - https://support.microsoft.com/Word/use-mail-merge-for-bulk-email-letters-labels-and-envelopes
2. n8n Docs - Basic LLM Chain node - https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.chainllm/
3. Mistral AI - Legal terms and data processing agreement - https://legal.mistral.ai/
4. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj