German AI voice agents for banking take customer calls under EU law, such as the AI Act and the Digital Operational Resilience Act (DORA). The fixed greeting names the bank and says the caller is talking to an AI. After a knowledge check, answers come from the core system, and self-hosted models keep the audio inside the bank.
Key Takeaways
- The AI notice and the bank's name go in the fixed greeting, in Sie form.
- Every hosted platform and model becomes an ICT provider on the bank's DORA register.
- Self-hosted Dograh with local models keeps call audio inside the bank's own tenant.
At a German bank, every outside service on an AI call becomes an entry in the bank's records. We cover which EU rules apply, a Sie greeting naming the bank and the AI, BaFin's three ways to host the model and German speech settings in Dograh. It belongs with our guide to banking voice agents.
Which rules apply to a German bank call
A German bank call runs under EU data, AI and resilience rules, and the US laws in our pillar guide do not apply.
The buyer is often small. In the Bundesbank's 2026 survey of smaller banks, 1,113 banks and savings banks took part, about 90% of all credit institutions in Germany. Each runs its own phone line and keeps its own list of outside providers.
The call data falls under the General Data Protection Regulation (GDPR), the DSGVO in German, and the EU AI Act adds an Article 50 notice. Since 2 August 2026, an AI system that talks directly with people must be designed so they are told they are dealing with an AI, at the latest at the first interaction, unless that is obvious. Because the Act treats credit scoring as a high-risk use, the agent never judges whether someone qualifies for credit.
Greet in Sie, name the bank and the AI
The first sentence carries the AI notice and the bank's name, so it is fixed text that the model never writes. In Dograh, that is the Greeting Text on the Start Call node, spoken word for word as the call connects. For a Sparkasse it might read: "Guten Tag, hier ist der KI-Assistent der Sparkasse Musterstadt. Dieses Gespräch wird aufgezeichnet. Wie kann ich Ihnen helfen?"
Callers can cut into a greeting after two words by default. Switch off Allow Greeting Interruption on the same node, so the AI notice plays in full on every call.
The agent speaks in the formal Sie throughout, and the prompt says so. German AI Voice Agents for Automotive explains why the recording notice matters under German criminal law, how to set the Sie form, and when a call counts as advertising.
After the greeting comes verification, and only then any account detail. A phone number finds the customer's record, but it does not prove who is calling. Before any balance or booking, the agent asks for an approved check, such as date of birth plus the last digits of the account number, and your core system answers pass or fail. The agent never asks for the full IBAN.
Some banks identify callers by their voice instead. A voiceprint used to identify someone counts as biometric data, a special category under the GDPR, and GDPR and Voice AI explains what that adds. A knowledge check against your own system avoids that question, and AI voice agents for KYC in banking covers the check in detail.
Open Source Alternative to Vapi / Retell
Self-hosted voice agent platform — no per-minute fees
dograh-hq/dograh
Star on GitHub
What BaFin's AI guidance compares
BaFin has written down the choice every German bank faces with an AI agent: where the language model runs. Its guidance on ICT risks when using AI is non-binding help for firms applying DORA, and it sets no supervisory expectations of its own.
It compares three set-ups for an assistant built on a language model. In the first, the model runs on premises, so all data flows stay inside the bank's own infrastructure. In the second, it runs in the bank's own cloud tenant, for example an open-source model the bank operates there. In the third, the model sits outside the bank's tenant, typically a provider's model called by API, and the guidance names the main risk plainly: data leaves the tenant and flows to the model provider.
A voice agent has more than one model, so the same question applies to speech-to-text and the voice as well. BaFin's three set-ups map onto two layers: the platform running the call flow, tools and telephony, and the speech and language models the bank picks for it. Self-hosted Dograh with local models falls in the first or second set-up. Hosted models on your own API keys fall in the third. Bringing your own keys changes whose contract covers the model, while the audio still reaches that provider.
The guidance also warns about lock-in. It suggests clarifying, before signing, which export formats a provider offers, naming Docker images and storage snapshots as examples. Dograh runs from Docker images, and its code is open under the BSD-2-Clause license, so the bank's own team can read the platform it runs.
DORA puts every vendor on the register
Under DORA, every outside service on the call becomes an entry in the bank's records. The Digital Operational Resilience Act requires a register of information covering all contracts for ICT services from third-party providers. For services that support critical or important functions, it also requires exit strategies with a transition period.
On a hosted voice platform, one customer call can add five entries: the platform, the speech-to-text provider, the language model provider, the voice provider and the phone carrier. Self-hosting the open-source platform generally takes it off that list as a hosted service, and local models take the model providers off too. The carrier stays on the register either way. So does the cloud host when the servers are not the bank's own, and a paid support contract with Dograh would be an entry as well.
Keep offers off these calls, because a product offer added to a service call can make it an advertising call under German competition law, which needs prior consent. Treat this post's points on DORA, the AI Act and advertising calls as general information, and confirm your bank's position with counsel.
Join the Dograh Community
Dograh is an OSS alternative to Vapi. Join our Slack community for queries, releases, best practices & community interactions.
German speech models in Dograh
German is well covered at both ends of the call. In the agent's Model tab, set the speech-to-text Language to German. If some customers switch to English, choose Multilingual (Auto-detect) and tick German and English under Language Hints, which work with Dograh's own models or Deepgram Flux. For a Swiss line, German (Switzerland) is a fixed language option on Deepgram Nova-3 and Dograh's own speech-to-text.
Dograh's own voices include German options, and its speech-to-text handles German. You can also bring your own keys for German voices from providers such as ElevenLabs or Azure Speech. Picking Azure's Germany West Central region sets where that provider processes the audio, though it is still a hosted service.
For BaFin's first or second set-up, Dograh's Local Models (Speaches) provider runs faster-whisper-large-v3 for German speech-to-text on your own servers. The local voice slot takes any compatible server that returns raw 24 kHz audio. For German, swap Kokoro, the default, for a German voice server and test it on a real call. German AI Voice Agents for Collections uses the same speech set-up for a different call.
Add terms such as Girokonto, Überweisung, Dispo and your product names to the Dictionary in the agent's Settings, so speech-to-text hears them. Have your core return dates and amounts written the way they should be spoken, such as "15. März" rather than "2026-03-15", then listen on a test call. QA summaries come out in English by default, so a German-speaking reviewer reads the transcripts of flagged calls.
Self-hosted, recordings and transcripts stay on your servers. The install also reports anonymous usage telemetry until your team sets ENABLE_TELEMETRY=false and removes POSTHOG_API_KEY, a step worth adding to the bank's hardening checklist. Running Dograh in-house carries no per-minute platform fee, and the bank pays instead for servers, model keys and phone minutes, while Dograh Cloud's platform fee is 1¢-a-minute.
Which call to hand over first
Start with questions that need no account, such as opening hours, fees and branch details from the Knowledge Base. Add balance and recent-transaction questions behind the knowledge check next. Fraud reports, blocked cards and disputes go to a person from the first day.
Transfer Call works on Twilio, Telnyx, Plivo, Cloudonix and Asterisk ARI. The advisor's phone rings first, and the customer joins only after the advisor picks up. On Twilio, both can hear a brief spoken introduction, in German if your voice speaks it. Before handing over, the agent writes the details to your case system, so the advisor opens the case already knowing why the customer called.
Test every agent on a real phone line before launch. Most phone calls reach the agent as 8 kHz audio, while a browser test runs at 16 kHz, so German compound words can sound different to the speech model.
Glossary
- Article 50 notice
- The EU AI Act duty, in force since 2 August 2026, to design AI systems that talk with people so those people are told they are dealing with an AI. On a bank call it belongs in the fixed Greeting Text, with greeting interruption switched off.
- Register of information (DORA)
- The record DORA requires of every contract for ICT services from outside providers. Each hosted platform and model provider on a voice call becomes an entry, with exit terms where the function is critical or important.
- Own tenant
- The bank's own isolated area in a cloud provider's infrastructure. BaFin's guidance compares a model running there with one outside it, where the main risk is that data leaves the tenant for the model provider.
- Knowledge check
- Verifying a caller with facts the bank already holds, such as date of birth and the last digits of the account number, checked by the bank's own system. Unlike a voiceprint, it uses no biometric data.
