EMAIL TRIAGE · USE CASE
Email triage automation: classify inbound flood, assign to client, prepare draft
IMAP watcher reads every mail, EU LLM classifies (client/invoice/query/newsletter/spam), RAG attaches client context, draft lands with the case handler. Dispatch only by hand.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What this is about
Email is still in 2026 the main channel between Swiss fiduciary firms, law offices, and SMEs. Per employee the inbox fills with 60 to 180 messages per day, ranging from a telecom collection letter to an urgent tax-deadline query from a key client. Sorting this manually costs a full-time case handler roughly 60 to 90 minutes per day – time that does not flow into consulting or processing.
Email triage automation means: an automatic pipeline reads every inbound mail, classifies it into fixed categories (client query, incoming invoice, newsletter, spam, other), assigns it to the right client (by sender or content), pulls relevant context from the client file via RAG, and places a reply draft in the responsible case handler's inbox. Sending happens only after human approval – automatic sending is here not just unwise but professionally problematic (Art. 321 SCC, professional secrecy).
The goal is not automating the reply. The goal is automating the preparation: sort, assign, research, draft. The case handler arrives in the morning, sees 80 mails – 35 of which carry a finished draft cutting writing time by 70%. The remaining 45 sit cleanly categorised and prioritised.
Why it matters
Three points make email triage in 2026 the best entry use-case for AI in a Swiss SME.
First: high leverage. Every person in a Swiss office spends 12 to 18 hours per week on email handling. Even 30% recovery through pre-sorting and drafts returns 4 to 5 hours per employee per week. In an 8-FTE office that is 30 to 40 hours per week.
Second: low risk. Unlike automated booking or direct client communication, a clean setup never sends anything automatically. The worst case is a poor draft that the human discards. There is no direct outside effect – the threshold for pilot acceptance in the office is low.
Third: good availability of building blocks. n8n has shipped for years a production-ready IMAP Trigger node with polling or IDLE mode. Microsoft Graph API gives the same access for Outlook 365 accounts. Mistral Small 3.1 (EU hosting, USD 0.20/1M in, USD 0.60/1M out as of May 2026) is enough in quality for classification and drafting. Per mail this yields roughly USD 0.0002 for classification and USD 0.001 to 0.002 for drafting – at 100 mails/day that is USD 0.20 to 0.40, about CHF 6 to 12 per month. Less than a lunch.
On regulation: client queries are personal data under Art. 5 lit. a revDSG. Sending them to a US LLM triggers Art. 16-18 FADP (cross-border transfer). With EU-hosted Mistral or Anthropic Claude via AWS eu-central-1 (zero-retention contract) the transfer stays within an acceptable frame. Precondition is a written DPA and an updated privacy notice.
How the pipeline works
The pipeline runs in five layers. It is deliberately kept simple – complex classifier chains often hurt reliability.
Layer 1 – IMAP watcher. An n8n workflow with the Email Trigger node listens on the inbox. IDLE mode (instant) is preferred over polling (every 1-5 min). Each mail gets a case ID and an audit-log entry. Attachments are extracted separately and optionally passed to OCR (see ai-belegerkennung-ocr) if PDF/image is detected.
Layer 2 – Classification. Mistral Small 3.1 (or Claude Haiku alternative) receives subject, sender and the first 1,000 chars of body. Output is a strict JSON: { category: "client|invoice|query|newsletter|spam|other", confidence: 0.0-1.0, sender_known_client: true|false, urgency: "high|medium|low", language: "de|fr|it|en" }. If confidence < 0.6 the case escalates automatically (layer 5).
Layer 3 – Client assignment. If sender_known_client = true, the client number is looked up via the email address in the Bexio/Klara/AbaConnect API. If false, a second LLM call searches the body for company names, client codes, or UIDs that map. If assignment stays uncertain the mail lands in an "unclear" folder and goes to a triage owner.
Layer 4 – RAG context enrichment. A Qdrant lookup retrieves from the client file the passages relevant to the mail content: latest correspondence, open items, contracts, VAT status. Top-5 passages feed into the draft prompt. Without RAG classification still works – but draft quality drops noticeably.
Layer 5 – Draft and approval. Mistral Small 3.1 writes a reply draft in the inbound language. Polite Sie-form, with concrete reference to the source material consulted. The draft appears as an Outlook draft (via Microsoft Graph) or Gmail draft (via Google Workspace API) in the responsible handler's inbox. They read, edit, decide – and only then send. The audit log records intake, classification, draft, approver, send time.
Edge cases. Pure newsletters or obvious spam produce no draft; mail is moved to the relevant folder. Incoming invoices route to OCR and produce a booking-prep sheet – a separate pipeline that this one only docks to. Multilingual mails (e.g. a French query from a Romandie client) are detected by language and the draft follows in the same idiom.
Pipeline in 6 steps
- 01Intake: n8n IMAP trigger (or Microsoft Graph for Outlook 365) processes every new mail, issues a case ID, writes the audit log.
- 02Classification: Mistral Small 3.1 (EU) classifies into 6 categories plus language, urgency, sender-known-client. Confidence < 0.6 escalates immediately.
- 03Client assignment: Bexio/Klara/Abacus API lookup by sender address, fallback LLM search in body for company name/UID.
- 04RAG lookup: Qdrant returns top-5 passages from the client file (latest correspondence, open items, contracts).
- 05Draft: LLM writes the reply in the appropriate language and register (Sie/Vous/Lei), with concrete reference to context. On uncertainty: note to handler without draft.
- 06Approval: draft as Outlook/Gmail draft in the inbox. Human reads, edits, sends manually. Audit log records approver and timestamp.
When to use
Email triage automation pays off from about 40 inbound mails per day per employee. Below that the setup overhead exceeds the benefit. From 80 mails per day per employee the project typically pays back in 3 to 5 months (realistic billable share 40-60% – see roi-rechner-ki-projekt).
Concrete constellations: a 6-FTE fiduciary with about 500 mails/day; a 4-lawyer law firm with 3 assistants and 700 mails/day; an 8-person SME sales team with 1,000 mails/day; a property-management firm with 12 case handlers and many rental queries.
Particularly well-suited: offices with clearly recurring query types (10 to 20 categories cover 80% of mail), offices with a well-maintained client file (RAG quality rises with good data), offices on Outlook 365 or Google Workspace (API integration is standard).
Less suited but possible: offices with a self-hosted mail server (IMAP works but setup takes longer), offices with a high share of hand-individual mails (lower hit rate because "recurring" does not dominate).
When not to use
Do not deploy without data-protection groundwork. Client mail contains personal data – without a DPA with the LLM provider and without an updated privacy notice the pilot violates revDSG (Art. 9 in conjunction with Art. 19 FADP). Groundwork takes 1 to 2 weeks.
Do not deploy if management expects mails to be sent automatically. That is not the use case – auto-send in client correspondence violates professional secrecy duties and is a reputation risk. Anyone wanting full auto-send is in the wrong use case (perhaps a marketing drip rather than triage).
Do not deploy without human-in-the-loop for at least the first 6 weeks. Even when the classifier becomes good: every classification must be checked in the ramp-up. Only when hit rate is stable above 90% over 4 weeks may manual spot-checks drop to, say, every 5th mail.
Do not deploy for clients under particularly strict professional secrecy (e.g. law firms with criminal cases) without local LLM routing. Here Ollama on own hardware is the clean solution; a cloud model has no place under attorney secrecy (BGFA Art. 13).
Trade-offs
STRENGTHS
- 60 to 90 minutes per day per employee saved – the highest-leverage use case in Swiss SMEs
- Low risk because nothing is sent automatically
- Building blocks have been production-ready for years (n8n IMAP, Mistral EU, Qdrant)
- Token cost under CHF 15/month at 100 mails/day with an EU model
WEAKNESSES
- Data-protection groundwork must be done before the pilot
- RAG quality depends on the client-file hygiene – bad data = bad drafts
- Encrypted mails (S/MIME, PGP) cannot be processed directly
- The first 6 weeks require active handler feedback, otherwise hit rate stagnates
FAQ
What happens to mails the classifier cannot assign?
They land in an "unclear" folder and with a triage owner. By design – better 5% unclear than 5% mis-assigned. In stable operation the unclear share settles at 3 to 8%. The error rate (mis-classified) is typically below 2%, since Mistral Small 3.1 is very robust for German and French.
How long does roll-out take in an 8-FTE office?
In practice 4 to 6 weeks. Week 1: data-protection groundwork + mail-type inventory. Week 2: IMAP trigger setup, classifier prompt, Bexio/Abacus integration. Week 3: shadow phase (classification runs but no drafts shown). Weeks 4-5: active phase with drafts, daily handler feedback, prompt tuning. Week 6: stable operation with spot-checks.
What happens to attachments?
Attachments are extracted separately by the IMAP trigger. PDFs and images optionally go to the OCR pipeline (see ai-belegerkennung-ocr); Office docs and ZIPs are mentioned in the draft but not opened. Attachments over 10MB are not piped through the LLM but only attached as metadata ("PDF, 12MB, invoice_2026-05.pdf") to the draft.
Can the pipeline handle encrypted mail (S/MIME, PGP)?
Not directly. Encrypted mails must be decrypted before classification – either by a local gateway (with the mailbox key) or by manual pre-check. In practice S/MIME is rare in Swiss SME context; most offices run unencrypted mail plus cloud confidentiality (Office 365 / Google Workspace, not E2E but TLS in transit).
Related topics
Sources
- n8n – Email Trigger (IMAP) node documentation · 2026-04
- Microsoft Graph API – Outlook Messages reference · 2026-03
- Mistral AI – La Plateforme pricing Mai 2026 (Small 3.1 USD 0.20/0.60) · 2026-05
- n8n.io – Workflow template: AI email automation with RAG · 2026-02
- EDÖB – Stellungnahme generative KI und Datenschutz, Update 2026 · 2026-01