What the AI sees. What it never does.

Email is the most sensitive data layer most people own. Before we touch a single message, you deserve a plain-language answer to five questions.

1. What does the AI analyze?

For every forwarded email, the assistant processes:

• The headers (from, to, subject, date)
• The body text (plain & HTML, stripped of trackers)
• Attachment metadata (filename, MIME type, size). Attachment contents are processed only if you enable "deep attachment analysis".

2. What does the AI actually do?

• Classifies importance (Critical / Important / Medium / Low / Spam)
• Generates a one-sentence summary and structured action items
• Extracts deadlines, amounts, and named entities
• Detects tone (e.g. polite-urgent, formal, friendly)
• Drafts suggested replies — never sent without your explicit click
• Triggers automations you defined

3. Where is the data processed?

• Storage and orchestration: AWS Frankfurt (EU)
• LLM inference: OpenAI & Anthropic EU residency endpoints with zero-retention contracts. Inputs and outputs are not logged on their side and are not used to train their models.
• Embeddings (vector search) are computed in-region and stored encrypted next to your account.

4. Do humans read my email?

No. Engineers cannot access message bodies. Production access is gated by hardware MFA and reads are logged in an immutable audit trail. The only humans who could possibly see content are:

• You (and people you explicitly share with)
• A support engineer — only if you paste the content into a ticket yourself
• A regulator with a valid court order — we will fight overbroad requests and notify you unless legally gagged

5. Do you train AI on my data?

No. Never. Not aggregated, not anonymized, not "just for evals". Our LLM provider contracts include zero-retention and no-training clauses. Our own classifiers are trained on synthetic data and a small set of public corpora — never customer email.

Confidence & the "Why" panel

Every AI decision in your inbox has a "Why" reasoning panel. Every importance label can be corrected with a click — and the correction only affects your account's per-user preferences, never a global model.

Known limitations

• The AI can mis-classify subtle social cues, sarcasm, or culturally specific phrasing.
• Suggested replies may be confidently wrong about facts. Always read before sending.
• Deadline extraction depends on date phrasing in your locale.
• Languages other than English, Swedish, German, French, Spanish, and Dutch are best-effort.

EU AI Act

Inbox Assistant is a limited-risk system under the EU AI Act. We meet the transparency obligations of Article 52: clear labelling of AI-generated content, summaries marked with a sparkle icon, and this page.