Private AI infrastructure
Your AI. Your rules. Your environment.
Post-train open-source models on company knowledge, run them on private infrastructure, and control every AI request. Keep sensitive data in-bounds and let only approved requests reach external providers.
AI traffic, governed
LiveSummarise Q2 revenue by region
Private modelDraft a launch email for a public release
Approved externalWhat is John Doe’s phone number?
Blocked · personal dataWhat is our parental-leave policy?
Internal knowledgeEmployees keep one AI workflow. You decide where every request runs.
A boundary you can explain
Keep knowledge, models and decisions inside your control.
Inside your environment
Post-trained open model
Adapted to your terminology and work.
Internal knowledge
Documents, wiki, CRM and approved sources.
Evaluation loop
Quality and safety measured continuously.
Audit log
Every request and routing decision recorded.
Cogent.io secure proxy
Policy decides where each request runs.
Approved external services
Approved external model
Only receives policy-approved, non-sensitive work.
Blocked request
Personal data or disallowed content does not pass.
A proven path to private AI
Discover
Inventory data, risks and useful work.
Post-train
Adapt open models to your knowledge.
Guard
Enforce policy and route intelligently.
Operate
Monitor, evaluate and improve.
Move from no AI to safe AI—without learning privacy the hard way.
The model interview process
Before a model starts work, it has to pass the interview.
A model can be technically capable and still be wrong for your organisation. We turn your ethics, values and reputation-sensitive boundaries into a repeatable acceptance test for every model you employ.
Our tested question bank is adapted to your values, policies, market and real-world edge cases.
Models that miss the agreed alignment threshold are remediated, restricted or rejected before they reach employees.
New models, major versions and material configuration changes are assessed before approval.
Learn your values
Interview leaders and translate business ethics, culture and reputation boundaries into testable expectations.
Build the interview
Select and adapt questions from our 100-question ethics bank, including realistic pressure tests and edge cases.
Interview the model
Score answers, refusal behaviour and consistency against an agreed alignment rubric.
Approve—or do not employ
Only models that meet the threshold enter service. Results become governance evidence for future re-tests.
The same offering, from another angle
Choose the story that best fits the people you need to bring with you.
Talk to us
Your AI. Your rules. Your environment.
Book a working sessionCommon questions
A short, practical FAQ
The legal answer always depends on the data, purpose and use case. These are the questions we use to start the right conversation.
Why can’t I just trust my AI vendor not to store data?
A strong vendor commitment is useful, but “not used for training” is only one part of the picture. Your organisation still needs to understand where a request is processed, how long logs and backups remain, which subprocessors can access it, whether data leaves the EEA, and how deletion, incidents and audits are handled. Under the GDPR, the controller remains accountable and must choose processors that provide sufficient guarantees. A local model and policy proxy reduce unnecessary exposure; they do not replace governance.
Are there really legal risks when processing personal information with AI in Europe?
Yes. Entering, analysing, storing or transmitting personal information through an AI system is still personal-data processing. The GDPR is technology-neutral, so you may need a lawful basis, clear purpose and transparency, data minimisation, appropriate security and retention controls. Special-category data and high-risk uses can require additional safeguards or an impact assessment. The EU AI Act can add further obligations depending on the system and how it is used. The risk is manageable—but it needs to be designed for.
Does running a model locally make us compliant?
Not automatically. Local operation can improve control over access, data residency, retention and auditability. You still need a defined purpose, appropriate permissions, model and data evaluation, user guidance, logging, incident processes and clear accountability. We treat the technical architecture and operating model as one programme.
Can employees still use approved external AI tools?
Yes. The proxy can keep one familiar workflow while applying policy to every request. Public, non-sensitive prompts can be routed to approved external tools; internal context and personal data can be routed to a private model or blocked. The policy is yours, and the decision can be logged.
Can a 100-question interview guarantee a model will never cause harm?
No single test can guarantee how a model will behave in every situation. The interview creates a measurable deployment gate aligned with your values, exposes important failure modes before employees encounter them, and records why a model was approved. We repeat it for new models, major versions and material configuration changes, and combine it with secure routing, monitoring and ongoing evaluation.
General information, not legal advice. Involve your DPO or legal counsel for your specific use case.