Privacy Policy

All Scopes — Complete List

📬 Email reading (gmail.readonly) is NOT requested — it is a Google-restricted scope. Reading support is planned pending security verification.

🔒 No restricted scopes are registered with our application.

How We Use Google Data — Limited Use Commitment

• Google user data is used ONLY to provide the specific feature you explicitly request in your chat message.
• We do NOT use Google data to serve advertising or for any marketing purpose.
• We do NOT allow humans to read your Gmail or Calendar data except to provide or improve the service, for security review, or as required by law.
• We do NOT transfer Google user data to third parties except as necessary to provide the service (e.g. AI interpretation), subject to the Limited Use requirements.
• We do NOT use Google user data to train AI or machine learning models.
• Google API access is read-on-demand only — no background scanning, no continuous sync, no automated access without your instruction.

Scope Justification

We request only gmail.send (a Google Sensitive scope — no security assessment required). We do not request gmail.readonly, gmail.modify, or any restricted Gmail scope. This means we have zero ability to read, search, list, delete, label, archive, or modify your inbox in any way — we can only send emails you explicitly instruct us to compose.

Account Information

When you create an account: name, company name, website, work email address, phone number (optional).

We do not store passwords. Authentication uses time-limited verification codes sent to your email.

Conversation History (AI Memory)

Your chat messages and summarised AI responses are stored in a vector database (Pinecone) to provide conversation context. When you send a message, we use your recent conversation (e.g. the last 30 messages) to interpret your request; we may apply an optional token cap and truncate older messages to keep context within limits. Important: we store only intent summaries for AI responses that retrieved tool data — raw email content, invoice details, CRM records, and third-party personal data are never persisted to our databases. Conversation data is automatically deleted after 7 days (weekly cleanup).

AI Usage Metrics

We log anonymised request metadata (tool used, action type, response time, token estimate, cost estimate) to operate and improve the service. This data is linked to your account for billing and support purposes.

Integration Connection Data

When you connect services (Gmail, Google Calendar, Asana, Jira, Notion, ClickUp, Xero, HubSpot), we store only encrypted OAuth tokens to maintain your authorised connection. We do not store the content of your data from these services.

Consent Records

We record the timestamp at which you accepted these Terms of Use and Privacy Policy, as required for GDPR Article 7 compliance.

🔒 Encryption in Transit

• All data between your browser and our servers uses TLS 1.3 / HTTPS encryption.
• All API calls to third-party services (Google, Jira, HubSpot, Xero, etc.) use encrypted HTTPS connections.
• No data is ever transmitted over unencrypted channels.

🔐 Encryption at Rest

• OAuth access tokens and refresh tokens are encrypted using AES-256-CBC before storage.
• Encryption keys are stored separately from encrypted data.
• Database connections use encrypted channels.

📧 Gmail Data Handling

• We do NOT read, store, or access any email content — incoming or outgoing — beyond what is necessary to send the email you compose.
• We have NO access to your inbox, drafts, sent items, or any stored email data.
• Only encrypted OAuth tokens are stored to maintain your Gmail send connection.
• When you disconnect Gmail, all OAuth tokens are permanently and immediately deleted.
• Gmail action available (only when you explicitly instruct): sending emails. We cannot read, search, list, reply, modify, archive, delete, label, or star emails.
• Scope used: gmail.send only. This is a Sensitive scope (not Restricted). No security assessment is required.
• Outbound emails sent via zizka.ai include a one-line AI disclosure footer by default (toggleable in settings).
• Email reading is coming soon, pending Google Restricted Scope security verification.

📅 Google Calendar Data Handling

• We do NOT store your calendar events or meeting details.
• Calendar data is fetched in real-time when you request it — no background syncing occurs.
• Event information is displayed in your chat and never saved to our databases.
• Scope used: calendar.events — minimum required for create, view, update, delete calendar events on your instruction.

🛡️ Access Control

• User-level data isolation: each user can only access their own data.
• OAuth tokens are user-specific and cannot be accessed by other users.
• Session tokens expire after inactivity.
• Rate limiting prevents abuse (60 requests/minute per user).
• All requests require authenticated sessions.

⏱️ Conversation Memory & 30-Day Auto-Deletion

• Conversation history is stored in Pinecone (vector database) with a 7-day expiry timestamp on every record.
• A weekly automated cleanup (Sunday 00:00) permanently deletes all conversation data older than 7 days.
• You can also delete your entire conversation history on demand at any time — see Your Rights below.
• AI responses containing fetched tool data (emails, invoices, contacts) are sanitised before storage — only an intent summary is saved, never the raw data.

EU AI Act Risk Classification

zizka.ai is classified as a Limited Risk AI system under the EU AI Act (Article 52 — transparency obligations apply). It does not fall under high-risk categories: no employment decisions, no credit scoring, no biometric surveillance, no critical infrastructure.

How Our AI Works — 6-Layer Architecture

• Layer 0 — Tool Intelligence Layer: Hardcoded expert knowledge base for all connected tools. Fires before the FAQ check. Provides instant concept explanations, onboarding guides, cross-tool comparisons, and guided workspace setup templates. Zero AI inference — all answers are static and deterministic.
• Layer 1 — FAQ Pre-Check: Regex-based hardcoded answers for common questions. AI never called.
• Layer 2 — Semantic Interpreter: GPT-4o-mini reads your intent and extracts structured parameters. Returns JSON only.
• Layer 3 — Guardrails: Validates and constrains actions; the AI cannot bypass this layer. In multi-step flows we may use try-then-ask (attempt then clarify if needed) rather than blocking on every missing parameter.
• Layer 4 — Decision Service: Our backend (not the AI) executes validated actions via tool APIs. The AI has zero direct access to your tools.
• Layer 5 — Response Sanitizer: Strips machine markers and formats responses as natural language.

Human Oversight (EU AI Act Article 14)

The AI never executes actions autonomously. Every action requires your explicit instruction. Multi-step actions require your explicit confirmation before execution. You can cancel any pending action at any time by saying "no" or "stop".

AI-Generated Email Disclosure (EU AI Act Article 52(3))

When you ask zizka.ai to send an email on your behalf, the outbound email includes the following footer by default: "This message was drafted with AI assistance (Zizka)." You can disable this in your account settings. This disclosure ensures recipients of AI-drafted communications are informed as required by Article 52(3).

No Autonomous or Background Processing

zizka.ai takes no action on your connected tools without your direct, real-time instruction. There is no background processing, no scheduled automation, no AI-initiated actions. If you are not in an active chat session, nothing happens.

Data Minimisation (GDPR Article 5(1)(c))

• Email list operations use Gmail's metadata-only format — full email bodies are never fetched during list/search operations.
• AI responses containing fetched tool data are sanitised before storing in our vector database — raw data from third parties is never persisted.
• Only data visible in your active chat session is passed to the AI model for interpretation.
• We do NOT use your data to train AI or machine learning models.

Incident Reporting

We maintain an internal incident log that captures tool failures and user-reported issues. If an AI response is problematic, you can report it directly from your chat interface. Incident records are used solely for service improvement and security review.

Automated Decision-Making (GDPR Article 22)

zizka.ai does not make automated decisions that produce legal or similarly significant effects on individuals. All actions on your connected tools are executed only in direct response to your explicit real-time instruction. You retain full control at all times.