DPIA support pack
1 Nature, scope, context & purpose
What the processing is: KYRA is an AI IT-support assistant that answers end-user questions across Microsoft Teams, Slack, a web widget, phone and SMS; retrieves relevant knowledge; can raise/read PSA tickets; and can perform gated Microsoft 365 actions after human approval. Scope: your staff and your clients’ staff/end-users. Context: a B2B service you deploy to your customers; you control which channels, integrations and AI provider are enabled.
2 Personal data & data subjects
- Data subjects: your and your clients’ staff and end-users.
- Data: identity & contact details; Microsoft 365 directory attributes; licence & device metadata; sign-in events; and support-request content (which may incidentally include other personal data a user types). KYRA does not require special-category data or credentials and instructs users not to share passwords, one-time codes or card details.
3 Recipients, sub-processors & transfers
Recipients are KYRA and the sub-processors on our sub-processor list (including the AI provider, and the telephony/voice providers when the phone channel is used). Some are US-based; those transfers are safeguarded by the UK IDTA and/or EU SCCs. Hosting and storage are in the UK (Azure UK South).
4 Retention
Conversation history is retained for a limited period (currently purged at 90 days) and other records (audit logs, ticket events, failed chats, notifications) per our schedule; on termination we delete or return data per the DPA. Confirm the periods that apply to you as part of your assessment.
5 Necessity & proportionality
KYRA processes only what is needed to answer a request, ground it in your knowledge base, and (where you enable it) raise a ticket or perform an approved action. Data minimisation, tenant isolation and access control limit exposure; the AI provider is controllable per tenant.
6 AI-specific considerations
- Accuracy & human oversight: AI answers can be wrong; they are framed as support suggestions, and privileged Microsoft 365 changes require explicit human approval — the model cannot execute them alone.
- No solely-automated decisions with legal or similarly significant effect are made about individuals.
- Prompt-injection resistance: knowledge-base content, tickets and messages are treated as untrusted data, not instructions.
- Training: support content is used to answer the request, not to train third-party foundation models (subject to each provider’s standard API terms as listed in the sub-processor list).
7 Risks & mitigations
| Risk | KYRA mitigation |
|---|---|
| Inaccurate AI answer acted upon | Answers framed as suggestions; privileged actions gated behind human approval; no solely-automated decisions |
| Excessive data exposure to the AI provider | Per-tenant provider control; data minimisation; no special-category data required |
| Unauthorised cross-tenant access | Tenant isolation; role-based access control; audit logging |
| International transfer risk (US providers) | UK IDTA / EU SCCs; UK-South hosting for storage |
| Credential / secret exposure | Secrets encrypted at rest (AES-GCM); users instructed never to share passwords/OTPs; logs record lengths/counts, not secrets |
| Prompt injection via untrusted content | KB/tickets/messages handled as data, not instructions; schema-validated actions |
8 What you still need to do
As the controller you should: confirm your lawful basis; complete your own risk assessment against your context; inform data subjects (privacy notice); and, if a high residual risk remains that you can’t mitigate, consult your supervisory authority (the ICO in the UK) before proceeding. We’re happy to answer reasonable questions to support your DPIA.
Contact
Kyra.bot Ltd — company number 17408470 (registered in England & Wales)
66 Paul Street, London, EC2A 4NA, United Kingdom
Data-protection contact: [email protected]