AI providers
AI Intelligence needs a large language model to run its analysis. You configure it once per organisation, using your own API key with a provider of your choice.
:::note Requires a Pro or Enterprise plan Configuring a provider is available on Pro and Enterprise. On Free the settings page is locked. See Plans and quotas. :::
Supported providers
| Provider | What you need |
|---|---|
| Anthropic (Claude) | An Anthropic API key (sk-ant-…) |
| OpenAI (GPT-4o) | An OpenAI API key (sk-…) |
| Custom / Self-hosted | A base URL and an API key |
| Hugging Face | An API key and a model ID |
Only one provider is configured per organisation.
Custom and self-hosted
The custom option works with any endpoint that speaks the OpenAI chat-completions API. You
supply a base URL (for example https://your-llm-host/v1) plus a key. This is the option for a
self-hosted model, an internal gateway, or a provider not listed above.
The base URL is validated when you save.
Hugging Face
Hugging Face requires a model ID in addition to the key, for example:
deepseek-ai/DeepSeek-V4.1-Flash:novita
The optional :provider suffix routes the request through a specific inference provider on
Hugging Face's router.
Configuring
An admin or owner configures this at Org settings → AI.
- Choose the provider.
- Paste the API key.
- For custom or Hugging Face, fill in the extra fields.
- Click Save & Validate.
Save & Validate really does validate
Unlike most settings forms, saving makes a live call to the provider to confirm the key works:
- Anthropic — a minimal request against a small model. Note that an HTTP 400 still counts as a successful validation, because it proves authentication passed and the key was accepted.
- OpenAI — lists available models.
- Custom / Hugging Face — calls the configured endpoint.
If validation times out rather than failing, the key is still saved, with a warning:
"Key saved but validation timed out. The key will be tested when you run your first analysis."
So a timeout is not a failure, but it is also not a confirmation. If your first analysis then fails, suspect the key.
Managing the key
| Action | Notes |
|---|---|
| Rotate key | Replaces the stored key with a new one |
| Remove | Deletes the configuration; AI Intelligence stops working for the organisation |
The stored key is shown only as a hint — the first ten characters followed by dots — so you can tell two keys apart without exposing either. The full key is encrypted at rest and cannot be displayed.
Validation state
The page shows when the key was last validated, or "Not yet validated" if it has never been confirmed by a successful call.
What data is sent to your provider
This is the disclosure shown in the product when you configure a provider. It is reproduced here exactly, because it is a commitment:
Your ticket titles and descriptions, and test case titles and descriptions are sent to your configured AI provider when running an analysis. No attachments, comments, or personal data are included.
| Sent | Not sent |
|---|---|
| Ticket titles and descriptions | Attachments |
| Test case titles and descriptions | Comments |
| Personal data |
:::warning Your tickets may still contain personal data The statement above describes what the application sends. It does not guarantee that the content is free of personal data — if someone wrote a customer's name or email address into a ticket description, that text is part of the description and will be sent.
If you work with personal data, review what your tickets typically contain before enabling this feature, and satisfy yourself that sending them to your chosen provider is appropriate for your obligations. :::
Because the key is yours, the content goes to your account with that provider and is governed by your agreement with them — including their retention and training practices. Choose accordingly.
Choosing a provider
| Consideration | Notes |
|---|---|
| Data residency | Where the provider processes and stores data matters for regulated workloads |
| Retention and training | Check whether the provider retains submitted content or trains on it |
| Model quality | Matching ticket text to test case text is a reasoning task; stronger models give better matches |
| Latency | A slow provider risks hitting the analysis timeout on large releases |
| Cost | You pay the provider directly; a large release can be many tickets' worth of tokens |
For sensitive content, a self-hosted model behind the custom option keeps data inside your own infrastructure, which is often the only viable choice under strict data-handling rules.
Limits
| Limit | Value |
|---|---|
| Configurations per organisation | 1 |
| Standard provider request timeout | 30 seconds |
| Custom and Hugging Face request timeout | 120 seconds |
| Whole-analysis timeout | 300 seconds soft, 360 seconds hard |
The longer timeout for custom and Hugging Face reflects that self-hosted and routed inference is usually slower.
Troubleshooting
| Symptom | Cause |
|---|---|
| "AI Intelligence requires a Pro or Enterprise plan." | The plan does not include the feature |
| Analysis fails immediately | The key is invalid — re-run Save & Validate |
| "Key saved but validation timed out" | The provider was slow; the key may still be fine |
| Analysis times out | Too many tickets for the provider's speed — analyse a smaller subset |
| Admin cannot find the page | The role must be admin or owner |