Governance
AI Model Hosting Governance
AI capabilities run on a routed, multi-provider architecture behind a single internal gateway, Minerva. This brief covers where Minerva and its providers are hosted, how customer configuration and data are scoped on shared serving infrastructure, and how AI or automated decision-making involvement can be surfaced on request.
This page answers
- How is the AI hosted and routed?
- Are AI models and data dedicated to each customer?
- Can a customer see which interactions involved AI or ADM?
- Is customer data used to train the AI models?
Routed multi-provider architecture
Rather than calling model providers directly from individual product features, all partner-facing AI is centralised behind Minerva, which routes requests to a pool of underlying providers: conversational agents, content and copy generation, recommendation scoring and other AI-assisted workflows.
Because every partner-facing AI request passes through one gateway, hosting location, provider choice and governance controls can be reasoned about and audited in one place rather than feature by feature.
The serving stack
- IsolatedPartner configurationPartner name, agent and knowledge base: the configuration and supporting data scoped to one customer account.
- SharedMinerva gatewayThe internal gateway for all partner-facing AI; routes each request to the appropriate provider in the pool.
- SharedProvider poolOpenAI, Azure OpenAI, Anthropic, Google and Amazon Bedrock.
- SharedAI application layerHosted in AWS eu-west-1; Bedrock inference is routed through the same region.
Layers and roles
| Layer | Role |
|---|---|
| Minerva | Internal gateway for all partner-facing AI; routes requests to the appropriate provider in the pool. |
| Provider pool | OpenAI, Azure OpenAI, Anthropic, Google and Amazon Bedrock. |
| AI application layer | Hosted in AWS eu-west-1. |
| Bedrock inference | Routed through AWS eu-west-1, where Bedrock is the serving provider. |
Shared serving, scoped configuration
- Partner name: the identifier tying configuration and data to a specific customer account.
- Agent: the specific AI agent or feature instance configured for that customer, such as a support agent or a shopping agent.
- Knowledge base: the customer-specific content the agent draws on, kept separate from other customers' knowledge bases.
- The model serving and supporting infrastructure is shared, not physically dedicated per customer. A requirement for fully dedicated models and data stores is a different architecture and should be flagged and discussed directly.
ADM and AI transparency
Where a customer needs to know which of its customer interactions involved automated decision-making or AI, or needs to support a request for human review of an AI-driven outcome, the product team provides this on request, drawing on the agent configuration, knowledge base and decisioning logic. A standing, self-serve inventory of AI models with purpose, inputs and decision impact is not yet a native capability.
Data use in model training
Customer data processed by the AI features is not used to train the underlying foundation models in the provider pool. A customer's data informs the responses generated by that customer's configured agents and knowledge base; it does not feed back into the shared models.
In TemasCRM
Every claim is compared with this product: what is live, what is partial, what is planned and what is not applicable by this product's design.
2 Live3 Not applicable
- One gateway routing to a pool of model providersNot applicable
No language model is connected in this product, so there is no gateway and no provider pool. The copilot panel was removed; agent and knowledge-base definitions are records of the workspace, and nothing runs them.
Agents - Configuration scoped by partner, agent and knowledge baseLive
An agent's configuration and its knowledge bases are records of the workspace, stored under the same row security as every other record and answered 404 to any other workspace; nothing is shared across workspaces. They are definitions only: nothing runs an agent or reads a knowledge base.
Agents - Which interactions involved AI, on requestNot applicable
No interaction involves AI here: no channel runs an agent and no model answers, so there is nothing to disclose. What can be shown is the record: every change to an agent or a knowledge base leaves an audit row.
Agents - No training on customer dataLive
Nothing is trained because no model is called; a customer's data never leaves the workspace for a provider.
- Hosting in AWS eu-west-1Not applicable
This product runs on the operator's own infrastructure; the hosting region is the operator's decision, not a provider region.