AI use note
Last reviewed 2026-09-01.
This note is a published draft pending legal review. The rules it describes are product enforcement and already apply.
- AI proposes, people decide: models draft, classify, translate and propose; consequential actions — sending, merging, dialling, deleting — require a person's explicit approval.
- AI is a licensed, metered capability with visible usage, budgets and hard caps; it is never an ambient always-on process.
- What the models see is limited to the fields the requesting feature needs, after PII classification and redaction rules run; provider calls are logged and attributable per workspace.
- No customer data is used to train foundation models; provider terms are selected for a no-training commitment before any engagement.
Model providers
| Provider | Model family | Engaged for customer data |
|---|---|---|
| Anthropic | Claude | Not engaged |
| Z.ai | GLM | Not engaged |
These are the model adapters the product ships. A provider only ever touches customer data after it appears on the subprocessor list — announced in the updates feed before engagement, per the enforced notified-before-engaged rule. Neither is engaged for customer data today; deployments run with AI capabilities gated and in test mode until a workspace licenses them.
Concrete model versions are selected per deployment through configuration and shown in the workspace's AI settings; they are not hard-coded claims in this note.
The standing rules
- No AI output sends, merges, deletes or dials anything without a person approving that specific action.
- Customer data is not used to train foundation models. Provider agreements must carry a no-training clause before the provider can be engaged.
- Every AI call is metered against a workspace budget with a hard cap; usage is visible to the workspace and to billing.
The full written policy pack (access management, acceptable use, incident response and this note) is available under NDA through the trust-centre document request flow.