Voicemail & call-note transcription routed to email or ticket

A plumber, a medical practice or a small office comes back to a list of voicemails that must be listened to one by one, often with poor audio and a number mumbled twice. A small flow transcribes each message, extracts caller, concern, callback number and urgency, and drops it into the right mailbox or ticket queue within minutes. This is not call-centre analytics (SPEECH-002) and not a voice bot: it only turns missed calls into actionable notes.

Problem & value

• Today: someone listens to the answering machine every morning, types notes into a mail or a paper list and forwards it. Urgent cases (a burst pipe, a cancelled appointment) wait until the right person has listened.
• AI value: the text is readable in seconds, urgent messages are flagged, nothing gets lost between people, and every call leaves a searchable trace in the ticket system.
• Honest limits: accuracy drops with dialects, noise and phone-line quality, and phone numbers and names are the most error-prone parts. The audio therefore always travels along. No sourced time-saving figures exist for this exact scenario; measure it yourself with a one-week baseline.

Variants

• Trades and workshops: customer name, address, fault, urgency to a dispatch mailbox or ticket.
• Practices, law firms, tax advisers: appointment requests and callback wishes; the content is highly sensitive, so processing stays local.
• Field staff: short spoken call notes after a visit, turned into a note in the CRM (see AUTO-025).
• Public sector: citizen office voicemails routed by topic to departments.
• Private use: answering-machine messages of a household or club mailbox.

How it works

1. voicemail (PBX / FRITZ!Box / Teams)
2. audio arrives as mail attachment or file in a folder
3. speech-to-text (language detection, German + dialect tolerance)
4. LLM step: summary, caller, callback number, topic, urgency (JSON)
5. rules: topic
6. recipient / queue; urgent
7. chat ping
8. mail or ticket with transcript + audio; unclear
9. human listens

• ASR: Whisper-class models are the baseline; the small and medium variants run on a CPU server, large variants on a modest GPU.
• LLM step: a small model (about 7-8B, local) or an EU-hosted API turns the raw transcript into fixed fields. Phone numbers are additionally checked with a regular expression.
• Human in the loop: the receiver sees transcript and audio, and corrects or deletes. The flow never replies to the caller by itself.

Data preparation

Which data: the voicemail audio, a routing table (who handles which topic, who covers absences) and optionally a customer list to recognise known callers by number.

Where it lives: in the PBX or router (FRITZ!Box answering machine, 3CX, sipgate, Teams Phone), as mail attachments in a shared mailbox, or on a Nextcloud/SharePoint folder.

Typical gaps: voicemails stay in the device and are never exported, the answering-machine greeting does not ask for name and number, no clear responsibility per topic, outdated customer phone numbers, and no deletion rule for old recordings.

Preparation steps: 1. Switch on mail dispatch in the PBX or router so that each message arrives as an audio file in one mailbox. 2. Update the greeting: name, number, concern, urgency. 3. Write the routing table as a simple spreadsheet (PREP-019 for categories). 4. Export the customer phone numbers for caller matching (PREP-005). 5. Collect 30-50 real, anonymised voicemails as a test set (PREP-018, PREP-022). 6. Define retention for audio and transcripts (PREP-025).

Tools: PBX/router settings, Excel or Nextcloud table for routing, a shared mailbox.

Rough share of effort: about 15 percent; the clean-up of greeting, mailbox and responsibilities matters more than the AI. See GUIDE-003.

Implementation paths

Which path for whom

• SMB — Path C: voicemails contain names, health or legal concerns and customer data; Whisper plus n8n runs on a small server (HW1/HW2) and the same stack later serves meeting notes and voice memos (Baukasten effect). Choose A instead if you already pay for Teams Phone and the built-in transcript is enough.
• Enterprise — Path B: a telephony platform and ticket system already exist; EU-region speech APIs plus own routing fit best, with the built-in Teams features for commodity cases.
• Developer — Path C: open-source ASR and workflow tools give full control and low running costs.
• Public — Path C: citizen concerns and procurement rules favour self-operated or German-cloud hosting.

Path A — Buy (off-the-shelf)

• Microsoft Teams Phone / Exchange Online: voicemail transcription is built in and is managed through Teams voicemail policies (e.g. EnableTranscription). Microsoft announced a transcription engine change from June 2026 with additional languages; check the supported-language table for German [1].
• Cloud telephony and PBX with voicemail-to-text: e.g. sipgate (German provider), 3CX, Placetel, Wildix; check for EU hosting, an AVV and whether transcripts are stored or used for training.
• Limit: the output is a transcript in a mail; routing by topic and tickets need extra flow steps.

Path B — Build on cloud APIs

• Speech APIs in EU regions: Azure AI Speech (EU regions), Google Cloud Speech-to-Text (EU location), or an EU-based provider; the OpenAI and Mistral APIs also offer transcription, so check hosting and AVV.
• Orchestration: Power Automate (trigger: new mail in shared mailbox), Zapier or Make, with an LLM step for the summary.
• Choose B when: you want no own server and accept an AVV with a cloud provider.

Path C — Self-host (open source)

• ASR: Whisper models via faster-whisper (CTranslate2-based, much faster than the original implementation) or whisper.cpp [2]; the medium model is usable on a CPU server, large-v3 on one workstation GPU.
• Summary and fields: Ollama with a small instruction model; schema-constrained JSON output.
• Workflow: n8n (self-hosted) reads the mailbox via IMAP or a Nextcloud folder, calls ASR and LLM, and creates mail or tickets (Zammad, Paperless-ngx tasks, OTOBO, Planner).
• Hardware: HW1 for a few dozen messages a day (a few minutes of processing per message on CPU is acceptable); HW2 for faster turnaround.
• Ops effort: about half a day per month for updates and a spot check.

Path D — Custom training

Rarely worth it. Only for strong dialects or domain vocabulary at high volume would a fine-tuned ASR model help; a vocabulary list in the prompt usually suffices.

State of the art (2026)

• Open Whisper-class models transcribe clean German speech well; large models cope better with noise and accents, but telephone audio (8 kHz) and dialects remain the weak spot.
• Built-in voicemail transcription in business telephony is now standard; the added value of own flows lies in extraction, routing and tickets.
• Failure modes: invented words in silent or noisy passages (hallucination), wrong digits in phone numbers, mixed-up names. Mitigation: audio always attached, number checked by pattern, low-confidence flag.

Shared resources & combinations

• Same stack (asr + llm-small + workflow-automation): SPEECH-001 meeting transcription, SPEECH-005 dictation, AUTO-025 voice memos; one Whisper server on HW1/HW2 serves all.
• Downstream: AUTO-013 contact capture into CRM, AUTO-011 request flows, AUTO-008 rows in a spreadsheet.
• Upgrade path: SPEECH-002 once volumes and analytics needs grow.

Effort, cost & KPIs

• Path A: hours to one day if the telephony product already includes it.
• Path B: 1-3 days for a flow with routing; API cost per minute of audio is small at office volumes (check current price lists).
• Path C: 2-5 days incl. server setup, test set and tuning; hardware reuses an existing server.
• KPIs: time from voicemail to responsible person, share of messages with correct number and topic, callbacks within the target time, lost or forgotten calls.

People & change

Involve early: management and the head of the front office as process owner; IT for PBX and mailbox; the DPO before callers' voices and content are processed. A short information note to staff who record call notes belongs in the plan (GUIDE-005, GUIDE-008).

Works council: depends. A flow that only handles external voicemails is less critical, but as soon as it logs who processes which call and when, or transcribes staff call notes, it is objectively suited to monitor behaviour under §87(1) no. 6 BetrVG. §90(1) no. 3 requires timely information about AI-supported work processes. Public bodies involve the Personalrat (GUIDE-004).

Data protection: callers did not agree to a transcript, so name the purpose in the greeting ("your message is recorded and transcribed to process your request") and cover it in the privacy notice; AVV (Art. 28 GDPR) with every cloud processor; a DPIA is usually not needed for plain voicemail handling but check with the DPO if health or legal content is common. Never transcribe live calls without a legal basis and a notice.

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; a 30-minute briefing on typical transcription errors and on checking the audio is enough (GUIDE-006).

Pilot: run for two weeks with one team in parallel to listening, count wrong numbers, then extend (GUIDE-009).

Typical concerns: surveillance of staff (use only for external messages, no per-person evaluation), errors (audio always attached), job loss (relief from listening and typing).

Risks, regulation & compliance

• EU AI Act: minimal risk; Art. 4 literacy applies. No emotion detection of callers, which would be a separate and sensitive topic. Art. 50 duties do not apply to the internal transcript.
• GDPR: voice recordings and transcripts are personal data and sometimes special categories (health); set a short retention, restrict access, prefer local processing (GUIDE-005, PREP-025).
• Telecommunications rules: recording of live conversations is different from voicemail; do not extend the flow to live calls without legal advice.
• Security: audio mails are an attack surface (phishing, prompt injection in the transcript); the LLM must never execute actions from transcript text (GUIDE-013).

Getting started

1. Count voicemails per week and note where they are stored today.
2. Switch on mail dispatch in the PBX or FRITZ!Box [3] and improve the greeting.
3. Test Whisper (Path C) or the built-in Teams transcript (Path A) with 30 real, anonymised messages.
4. Build the flow with a routing table and urgency flag; always attach the audio.
5. Inform the DPO and works council, document retention, run a two-week pilot.

Sources

1. Microsoft Message Center MC1338827 — Teams voicemail transcription service migration and new languages — https://mc.merill.net/message/MC1338827
2. faster-whisper (SYSTRAN, CTranslate2-based Whisper inference) — https://github.com/SYSTRAN/faster-whisper
3. GIGA — FRITZ!Box Anrufbeantworter einrichten und bedienen — https://www.giga.de/unternehmen/avm/tipps/fritzbox-anrufbeantworter-einrichten-und-bedienen/
4. Microsoft Learn — Change the default language for greetings and emails (Cloud Voicemail) — https://learn.microsoft.com/en-us/microsoftteams/change-the-default-language-for-greetings-and-emails
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj