AI Transcription That Keeps Your Data in Europe

Note: This article describes Mandraki’s architecture and design. Some features discussed are being rolled out progressively and may not yet be available in all plans.

AI-powered transcription and meeting summarisation have become expected features in enterprise collaboration tools. They save time, improve accessibility, and create searchable records of discussions that would otherwise exist only in participants’ memories.

But for European organisations, these features come with a serious question: where is the audio being sent, who is processing it, and what happens to it afterwards?

With most major collaboration platforms, the answer is uncomfortable. Audio is streamed to data centres operated by US companies, processed by models hosted on US-controlled infrastructure, and retained under terms governed by US law. For a European hospital discussing patient care, a bank reviewing financial strategy, or a government ministry coordinating policy, this is not an abstract regulatory concern — it is a real data protection risk.

Mandraki takes a fundamentally different approach.

Processing within EU jurisdiction

Every component of Mandraki’s AI pipeline runs on European hyperscale infrastructure within the European Union. The audio capture, the speech-to-text processing, the summarisation model, and the storage of results all occur on servers subject to EU law, operated by a European-owned cloud provider, with no US parent company or jurisdiction exposure.

This is not a matter of selecting an EU region within a US cloud provider’s console. Our infrastructure provider is a European company, incorporated under EU law, with no foreign ownership structure that could create indirect jurisdictional exposure. When your audio is processed by Mandraki’s AI features, it stays within the same sovereign boundary as the rest of your data.

How transcription works

When an organisation enables AI transcription for a call, the audio stream is captured by our Selective Forwarding Unit and sent to the transcription service running on the same European hyperscale infrastructure. The speech-to-text model processes the audio in near real-time, generating a timestamped transcript with speaker attribution.

The transcript is then encrypted using the organisation’s Data Encryption Key — the same three-layer envelope encryption that protects all other data in Mandraki — and stored alongside the call record. For BYOK organisations, the transcript is encrypted with the customer-controlled key hierarchy, ensuring full cryptographic control.

After processing, the raw audio buffer is discarded from memory. It is not written to disk, not retained for model training, and not accessible to Mandraki staff.

Summarisation and smart features

Beyond transcription, Mandraki offers AI-powered meeting summarisation. At the end of a call (or on demand during a call), the summarisation model processes the transcript to generate a structured summary: key discussion points, decisions made, action items identified, and questions raised.

These summaries are designed to be useful without being reductive. They preserve the substance of a discussion while making it scannable. For a thirty-minute call, a typical summary is two to three paragraphs — enough to remind a participant of the key points, or to brief a colleague who could not attend.

Like transcripts, summaries are encrypted at rest with the organisation’s key hierarchy and subject to the same data retention policies.

The mutual exclusion with E2EE

We believe in being transparent about architectural constraints. AI transcription and end-to-end encryption are mutually exclusive in Mandraki. This is not a limitation we can engineer away — it is a fundamental property of E2EE.

End-to-end encryption means the server cannot access plaintext content. AI transcription requires the server to process plaintext audio. These two requirements are logically incompatible.

Mandraki enforces this at the architectural level, not just through policy. A call with end-to-end encryption enabled cannot have AI features turned on, and vice versa. The validateAiE2eeMutualExclusion guard checks this constraint before any AI operation is permitted.

Organisations choose their preference at the call or channel level. A company might use E2EE for board meetings and legal discussions while enabling AI transcription for general team standups. The choice is granular and explicit.

Consent and governance

AI features in Mandraki are governed at multiple levels, reflecting the principle that organisations and individuals should retain control over how their communications are processed.

At the organisation level, administrators set an AI policy: disabled (no AI features available, E2EE available), opt-in (AI available but requires per-call consent from participants), or enabled (AI on by default). They can also toggle individual features: transcription, summarisation, smart replies, and recording analysis.

At the call level, when the organisation policy is set to opt-in, participants are prompted to consent before AI features activate. Consent is logged with a timestamp and the specific features consented to. If any participant declines, the call proceeds without AI features.

This consent model is designed to satisfy GDPR requirements for lawful processing, particularly in jurisdictions where employee monitoring is subject to strict rules. The consent log provides an auditable record that can be produced to data protection authorities if required.

Data retention

Mandraki supports three AI data retention modes, configurable at the organisation level.

Transient. AI-generated content (transcripts, summaries) exists in memory during processing and is encrypted and stored, but the raw input (audio buffers) is never persisted. This is the only mode available for BYOK organisations.

Session. AI-generated content is retained for the duration of the call and a configurable grace period afterward, then automatically deleted.

Persistent. AI-generated content is retained according to the organisation’s data retention policy, encrypted at rest, and included in data export.

In all modes, raw audio is never retained beyond the processing window. Mandraki does not use customer audio or transcripts for model training or improvement. The AI models are not fine-tuned on customer data.

A European approach to AI in the workplace

The European AI Act establishes risk categories and transparency requirements for AI systems. While collaboration AI features like transcription and summarisation generally fall into lower risk categories, the principles of transparency, consent, and data minimisation are central to Mandraki’s approach.

We believe that AI features should make work easier without requiring organisations to compromise on data sovereignty or employee privacy. Processing within EU jurisdiction, encrypting outputs with customer-controlled keys, enforcing consent, and providing granular retention controls are how we deliver on that belief.