All three supported providers produce good observation drafts. Choose by the processing location, retention arrangements and contractual terms your organisation needs. This guide explains our managed accounts, what changes with BYOK, and the separate path used for semantic matching.
Data agreements, retention and model memory
A data processing agreement (DPA) governs how a provider handles data on our behalf. Retention describes how long the provider keeps inputs and outputs. Zero data retention (ZDR) removes the normal content-retention window for requests covered by that agreement. A DPA alone does not establish zero retention.
Model memory is another distinction. Our generative requests are stateless: the model does not remember one request as context for the next. That tells you how an answer is generated, not whether the provider retains request data under its terms.
Choose the processing terms that fit
Advanced AI is included on Pay-per-use and Pro. We hold DPAs with all three managed providers, and the API requests we make are not used to train their models. The processing location and retention arrangements differ:
| Provider | Processing | Content retention |
|---|---|---|
| Anthropic | United States | ZDR enabled |
| Mistral | European Union | ZDR enabled |
| OpenAI | United States | Standard provider retention; no ZDR on our account |
Anthropic is the default and carries ZDR on our managed account. Mistral adds EU processing with ZDR. OpenAI is available for teams that prefer that provider and can accept its standard retention arrangements. These are the terms for our managed accounts; they do not automatically apply to your own provider account.
An organisation administrator can change the managed provider in the AI settings. All three support the same product workflows. Free uses an open-weights model with 120 billion parameters, delivered by Cloudflare. The managed-account agreements in this table do not apply to that service.
Temporary processing storage
We store assembled prompts only for the duration of processing, normally measured in seconds, and clear them when processing is finalised. Encrypted drafts and results can remain briefly so an interrupted workflow can be recovered. Saved observations are part of your register, not a memory held by the model.
Those are separate from the provider's retention policy. Our AI architecture guide explains how we limit context and validate model output.
What BYOK changes
With Bring Your Own Key, generative requests use your OpenAI, Anthropic or Mistral account. Your agreement, processing settings and retention arrangements apply. Confirm them with your provider before submitting sensitive findings; our managed-account ZDR agreements do not transfer to your key.
Semantic matching uses Mistral
When semantic matching is enabled, observation text is sent to our platform-managed Mistral embeddings service under our DPA and ZDR agreement. This applies on every plan and regardless of the generative provider, including BYOK. The resulting vectors are encrypted in our EU-hosted database.
Saving or replacing a BYOK key automatically turns semantic matching off and deletes its stored vectors. To use it again, an administrator must enable it manually in AI settings and confirm the separate Mistral processing. Leave it off if your findings must go only to your own provider account.
Keyword matching remains available without semantic matching. See how duplicate matching works for the difference between the two.