Integrations
Mistral AI
How Mistral AI's language models, including its open-weight models, can be used in your deployment, where they run and what they are allowed to see.
On this page
Mistral AI's models are among the language models Telonic can use in your deployment. The language model (the AI component that interprets the conversation and composes the reply) understands what the customer is asking and decides the reply, within the agent's limits. The model is chosen per deployment, with you, and can be changed later. Which providers and models are available in your region is confirmed during implementation, and named in your agreement.
What the model does in your deployment
| Task | What it means for you | Example |
|---|---|---|
| Understand the request | Works out what the policyholder wants and picks out the details that matter | A caller describing a burst pipe is understood to be reporting a new home claim |
| Decide the reply, within limits | Chooses what to say and do from the actions your team permits | Explains the policy's excess but does not comment on whether the claim will be paid |
| Express facts from your systems | Decides the wording and order. Figures such as amounts and dates are inserted directly from the system record | The excess read exactly as the policy record states it |
| Ask when something is unclear | Asks a question back rather than guessing | "Was the water coming from inside your flat, or from the one above?" |
How it connects
All processing for your deployment runs in the region you choose, unless you explicitly choose a language model provider outside it. If you do, personal information is replaced with placeholders before any text leaves your deployment, and the real values are restored in the reply inside your deployment.
Telonic's contract with every provider prohibits using your data to train or improve their models, and zero data retention arrangements (the provider keeps no copy of what it processes) are used wherever the provider offers them. A new model is adopted only when it performs at least as well on the test set for your industry, and any change of model or provider needs your organisation's approval before release. If a provider has an outage, failover switches only to another approved provider in your chosen region, tested in advance and configured during implementation.
Running on your own premises
Mistral AI offers some of its models as open-weight models (models whose files are released for others to run on their own infrastructure). Because the model files are published, these models can run on your own premises, on hardware you control. On-premises deployments run open-weight language models and self-hosted speech models, and the computing capacity they need is sized with you during implementation.
What the model can see
| Access | Needed for | Default |
|---|---|---|
| The conversation, the facts retrieved for the reply and the agent's instructions | Composing each reply | Given for each reply |
| Personal information, when the model runs outside your region | Not needed | Replaced with placeholders before the text leaves your deployment |
| Direct access to your systems | Not needed | Not given. Lookups and actions run through Telonic's software, within the permissions your team has granted |
| Your data, for training the provider's models | Not permitted | Prohibited by contract |
Whichever model you choose, the agent can only take actions your team has granted. See Policies as limits.
In practice
A Saudi insurer requires every system that handles policyholder conversations to run in its own data centre.
- Telonic and the insurer's infrastructure team size the computing capacity for an on-premises deployment.
- An open-weight Mistral model is chosen and runs on the insurer's own hardware, alongside self-hosted speech models.
- The model is tested on the insurer's claims conversations before the agent starts handling real customers.
- Only operational health signals, with no customer data, are sent to Telonic, and the insurer can switch them off.
Related
- Language modelsIntegrations
- Models and providersAgents
- Hosting and data residencySecurity and data protection
- Personal information redactionSecurity and data protection
- Testing before every releaseGovernance and control
Product names and logos are trademarks of their owners. Their mention shows systems Telonic connects to and does not imply partnership or endorsement.