Den här texten finns bara på engelska.
AI Transparency Disclosure
Deep Thought Cloud
- Last Updated: October 7, 2026
- Version: 1.2
Provider
Dolutions AB Org.nr 556981-0293 Box 150, 101 22 Stockholm, Sweden VAT: SE556981029301
Contact: legal@deep-thought.cloud
1. AI-Powered Service
Deep Thought Cloud uses artificial intelligence to process your content. This document explains how AI is used in our service, in accordance with the EU AI Act transparency requirements.
2. AI Capabilities
Our service uses AI for the following functions:
| Function | Description |
|---|---|
| Transcription | Converting speech in audio/video files to text |
| Speaker Recognition | Naming or suggesting who is speaking by comparing voices with voice profiles from earlier recordings in your account, on our own hardware (see our Privacy Policy, §5) |
| Meeting speaker suggestions | In meetings recorded by the meeting assistant, suggesting who is speaking from the participant list and the meeting platform’s speech timeline; a participant the platform identifies, and whom you have confirmed in three meetings, is named automatically in later meetings — never from the voice alone (see our Privacy Policy, §6) |
| Summarization | Generating concise summaries of transcribed content |
| Entity Extraction | Identifying people, organizations, locations, and topics |
| Keyword Extraction | Identifying key terms and concepts |
| Auto-Tagging | Automatically suggesting tags based on content |
| Content Analysis | Analyzing content type, language, and context |
3. AI Models and Processors Used
Processing happens in two stages with different guarantees. This section lists every service that may process your content.
Stage 1 — Speech-to-text and speaker identification (always our own hardware)
| Model | Purpose | Runs on | Jurisdiction |
|---|---|---|---|
| KB-Whisper | Swedish speech recognition | Our own GPUs | Sweden |
| OpenAI Whisper (local deployment) | Multilingual speech recognition | Our own GPUs | Sweden |
| pyannote (local deployment) | Speaker identification and separation | Our own GPUs | Sweden |
Audio you record or upload is never sent to a third party. No account setting changes this. The one case where a third party handles audio is the meeting assistant (below): Recall.ai records the meeting and passes the audio to us, and the transcription runs on our own hardware as in the table above.
Meeting capture — the meeting assistant
| Processor | Purpose | Jurisdiction | AI processing |
|---|---|---|---|
| Recall.ai (Hyperdoc Inc., USA) | Joins Zoom, Microsoft Teams and Google Meet meetings you send the assistant to, records the meeting’s audio, and passes the participant list and speech timeline from the meeting platform; reads connected calendars to decide which meetings it joins | EU region (Frankfurt, Germany) for the recording; a US company, which may access the data from the United States under the EU Standard Contractual Clauses | None for us — Recall.ai’s transcription is switched off; the recording is erased at Recall.ai once our copy is verified, and at the latest after 7 days |
The assistant joins under the name “
Stage 2 — Text analysis (depends on your account settings)
The transcribed text is analysed to produce titles, summaries, entities, keywords and insights. The processor depends on your settings and on availability, in this order:
| Order | Processor | Model | Jurisdiction | When it is used |
|---|---|---|---|---|
| 1 | Your own provider (BYOK) | Your choice | Per your contract | Only if you have configured an API key — then the only external processor for your account (§4.1); rows 2-4 apply to accounts without their own key |
| 2 | Berget.ai | gemma-4-31B-it |
Sweden (EU/Nordic provider, Sweden-resident inference) | The default for Premium and Enterprise accounts without their own key |
| 3 | Our own hardware (Ollama) | qwen3, gpt-oss, gemma4 |
Sweden, our own GPUs | When Berget.ai is unavailable, always in local-only mode, and for all AI on Free accounts (the title) |
| 4 | OpenRouter (Anthropic Claude Haiku) | claude-haiku |
United States | Last resort only — when both the above are unavailable |
Additional processors that may appear in specific features: system Anthropic (United States) for certain synthesis tasks.
Berget.ai is a third party. It is a Swedish company performing inference in Sweden, which is why we chose it as the default — but your text does leave our systems to reach it. Any statement that no data is sent externally applies only to Free accounts and to local-only mode (§4.2).
Images and PDF pages (Premium and Enterprise) are read by a vision model at Berget.ai (gemma-4-31B-it, Sweden), with OpenRouter as the last resort when Berget.ai is unavailable, or by the provider of your own API key (§4.1). Free accounts do not process images or PDFs.
4. Controlling where your content is processed
4.1 BYOK (Bring Your Own Key)
If you configure your own API key from OpenAI, Anthropic, Berget.ai or OpenRouter:
- The key you choose as primary is used for all text analysis and for images and PDF pages, under your own contract
- Processing occurs according to their privacy policy
- This is optional and user-initiated
With your own key, your provider is the only external one. If your key stops working — refused, out of quota, or the provider unreachable — processing waits and you are told by e-mail and in the Web App; it is not sent to Berget.ai, OpenRouter or any other provider in §3 instead. It resumes once the key works again.
4.2 Local-only mode
Local-only mode (local_processing_only) prevents all external inference for transcription and text analysis and for images and PDF pages. With it enabled:
- All text analysis runs on our own hardware
- Cloud transcription spillover is disabled
- Cloud-dependent integrations, including meeting-bot capture, are unavailable and will refuse to start rather than proceed
The restriction is enforced at five independent points in our system. Three of them fail closed: if the privacy check itself cannot complete, processing is treated as local-only rather than being allowed to continue. A failure in our own systems cannot cause your content to be sent externally.
Images and PDF pages wait in local-only mode. They are stored but not read: no vision model is called, the file shows that it waits because your account is local-only, and it is processed if you turn local-only off.
Trade-off: local-only mode uses smaller models than the default. In our own testing (2026-09-10) this cost processing speed rather than accuracy — the same meeting produced the same people, organisations and decisions, but took substantially longer. It is available on request — contact support@deep-thought.cloud.
5. AI Output Accuracy
Important: AI-generated outputs may contain errors.
- Transcriptions may have inaccuracies, especially with accents, technical terms, or poor audio quality
- Summaries are machine-generated interpretations, not human-verified
- Entity extraction may miss or misidentify items
- Auto-tagging suggestions are algorithmic recommendations
We recommend human review for:
- Legal or contractual content
- Medical or healthcare information
- Financial decisions
- Any critical business use
6. Human Oversight
- AI outputs are provided as-is without human verification
- Speaker names suggested from a meeting’s participants are suggestions until you confirm them; automatic naming applies only to a participant you have confirmed in three meetings, and you can correct any name
- You can edit, correct, or delete any AI-generated content
- Feedback mechanisms allow you to improve accuracy over time
- Final decisions should always involve human judgment
7. Data Usage
- Your content is NOT used to train our AI models
- Processing is performed only to deliver the requested service
- Audio files are kept with the recording on our own servers in Stockholm so you can play them back (see our Privacy Policy, §4 and §7)
- Meeting recordings made by the meeting assistant are kept the same way; Recall.ai’s copy is erased as described in our Privacy Policy, §6
- See our Privacy Policy for full data handling details
8. AI Risk Classification
Under the EU AI Act, our AI system is classified as limited risk, requiring transparency obligations. We provide this disclosure to ensure you are informed about AI usage in our service.
9. Your Rights
You have the right to:
- Know when AI is processing your content (this disclosure)
- Access AI-generated outputs
- Correct or delete AI outputs
- Opt out of optional AI features
- Request human review where applicable
10. Contact
Questions about AI usage: legal@deep-thought.cloud
This AI Transparency Disclosure is provided in compliance with EU AI Act transparency requirements.
Changes in version 1.2 (October 7, 2026)
- §2: meeting speaker suggestions added.
- §3: the meeting assistant and Recall.ai (Hyperdoc Inc.), its processor for meeting capture — the recording stored in its EU region in Frankfurt, access from the United States under the EU Standard Contractual Clauses, audio only and no transcript at Recall.ai; the statement that audio is never sent to a third party now covers audio you record or upload, with the meeting assistant as the one exception.
- §6: how meeting speaker suggestions are confirmed.
- §7: meeting recordings; the Privacy Policy reference updated to its renumbered sections.