Control over your AI traffic and data
AI rarely fails a security review on capability. It fails on the questions nobody can answer precisely: which models, whose data, held where, evidenced how. This page answers those in the order they are usually asked.
The five questions your security review will ask
These are the questions that stall AI procurement. Our answers are below in the order they usually arrive.
Which models can our staff actually reach?
Each team and data class is pinned to approved models and regions. The check runs before the request leaves your boundary, so clinical or patient data can be restricted to UK-hosted or internal models only.
What stops sensitive data being sent in the first place?
A pre-call guardrail checks every prompt for PII and PHI, then allows it, blocks it or reroutes it to an internal model. Nothing leaves the boundary while that decision is being made.
What do you store, and for how long?
We record the token counts the provider returns, along with who made the request and which model it went to. We do not store the prompt or the response text. Storing that content is off by default and opt-in per policy, so in the default configuration we hold none of it.
What can we hand to an auditor?
Each request is recorded with the time, the application and team behind it, the data class, the model requested and the outcome. The record is exportable to support compliance reviews.
Where does the control layer itself run?
In the UK, in the EU or inside your own cloud. UK and EU are treated as separate jurisdictions, each with its own policy.
A record of every request, exportable for compliance review
| Time | Application | Team | Data class | Requested model | Outcome |
|---|---|---|---|---|---|
| 09:14 | patient-record-summariser | Healthcare Delivery | Patient data | gpt-5 | Blocked → internal-llama |
| 09:02 | care-plan-assistant | Healthcare Delivery | Patient data | internal-llama | Allowed |
| 08:57 | hr-policy-chat | Development | Confidential | claude-sonnet-5 | Flagged sensitive-data policy |
| 08:51 | research-notebooks | LLM Research | Internal | claude-opus-4-8 | Allowed |
| 08:39 | triage-copilot | Production | Patient data | claude-sonnet-5 | Blocked → internal-llama |
Every model carries its own ceiling
Each model in the registry is recorded with the region it runs in and the highest data class it may receive. That pairing is what the gateway checks against, which is why a policy decision does not depend on anyone remembering where a given model is hosted.
Sovereign by design
Choose the deployment pattern that matches your regulatory position. The policy and ledger model stay the same in each case.
UK-hosted SaaS
Runs in our UK data-centre environment. EU hosting available as a separate jurisdiction. Fastest to onboard.
Customer cloud / BYOC
Bring your own cloud: deployed into your own tenancy, for outsourcing due-diligence or data-residency requirements.
Private / on-premise
For the most sensitive environments, self-hosted models sit fully inside your own boundary.
Happy to talk this through
A conversation is usually quicker than a questionnaire. Bring whatever questions you have about hosting, data or access and we will go through them with you.