Identity and security
Secure gateway for AI models
Every AI request goes through a shared gateway with masking, policies, quotas and cost metadata.
Fictional scenario. An example situation that illustrates the approach. It is not a client project and describes no measured results.
Business problem
A fictional company has several teams using different AI APIs with personal keys. There is no control over data or costs.
Assumptions
- Teams can change the API address.
- There is an AI usage policy.
- There is a budget for AI usage.
Constraints
- No personal data to models without a basis.
- Teams hold no keys.
- No content stored by default.
Reference architecture
Components by stage — an illustration, not a running system and not a certified design.
01Clients
- Internal applications
- Assistants
02Gateway
- Authentication
- Masking
- Policies and quotas
03Providers
- Model adapters
- Fallback provider
04Control
- Usage metadata
- Cost per team
The diagram as a list
| Stage | Components |
|---|---|
| Clients | Internal applications, Assistants |
| Gateway | Authentication, Masking, Policies and quotas |
| Providers | Model adapters, Fallback provider |
| Control | Usage metadata, Cost per team |
Interfaces
| From | To | Via |
|---|---|---|
| Applications | Gateway | Internal API with identity |
| Gateway | Providers | Adapter with a central key |
| Gateway | Counters | Metadata without content |
Data flow
- The application authenticates to the gateway.
- Personal data is masked.
- Policies check the model and the quota.
- The request goes to the provider.
- Usage is recorded without content.
Security boundaries
- Provider keys exist only in the gateway.
- Masking before sending.
- Forbidden models and data are refused.
Where a person approves
- A new model or provider.
- Masking exceptions.
Delivery phases
01 · Policy
Rules and data classes.
02 · Gateway
Authentication and masking.
03 · Migration
Teams onto the gateway.
04 · Control
Quotas and cost reviews.
Limitations
- Masking is not perfect.
- It adds latency to requests.
Trade-offs
- A central gateway — control vs a single point of failure.
- Strict policies — less flexibility for teams.
Related services
Related assessments
Integration patterns used
Workshops and assessments
Facing something similar?
We start with an assessment of where you are and propose a realistic first phase.