AI for sensitive data, Data cannot leave
Contracts, regulation, or plain caution rule out sending records to a third-party model.
Private models and governance
AI for sensitive data, on infrastructure you control.
We deploy open-weight models on your servers or in a region you approve, and put access policies, audit trails, and testing around every assistant and agent.
The problem
Why sensitive work stays manual.
No view of what AI did
Without records, no one can answer what an assistant read or changed.
No rules for who may use it
Access is granted ad hoc, and revoked late or never.
Our approach
What we build.
Self-hosted models
Open-weight models deployed on your hardware or a cloud region you choose.
Access policies
Who may connect which assistant, and what each may read or change.
Audit trails
A record of every action an assistant or agent took, kept as long as you need.
Testing before release
Agents are tested against misuse and bad inputs before they reach your staff or customers.
How Desk handles AI access
Policy, consent, and approval, built in.
In Desk, organisations choose which assistants may connect, connections expire and can be revoked, and writes need approval unless policy says otherwise.
See DeskHow it works
Three steps, each one reviewable.
Set the rules
We agree where data may go, who may use what, and what must be logged.
Deploy and connect
We stand up the models and connectors inside those rules.
Operate
We monitor, update models, and review access with you over time.
Privacy and control
Your data, your rules.
- Your data stays in systems you control, or in a region and provider you approve.
- Assistants and agents act with the access of the person using them, never more.
- Anything that sends, pays, or deletes waits for a person to approve it.
- Every action is recorded, so you can see what was done and by whom.
Questions
What teams ask first.
Are open-weight models good enough?
For many document, drafting, and classification tasks, yes. We test against your own tasks before recommending one.
Can we mix private and hosted models?
Yes. Sensitive work can stay on private models while other tasks use a hosted provider you approve.
Who runs it after launch?
Your team, us under a support agreement, or both. Everything is documented for handover.
Start a conversation
Private AI for the work that cannot leave your servers.
An engineer replies within 24 hours with how we would approach it.