The control plane for enterprise AI.

One gateway. Any model. Every request inspected, every policy enforced, every decision logged. Operated in Europe.

Book a demo See how Guard works

what your team sees what you see

Voban

Mistral Large EU

Message Voban...

Guard live

12,847requests

342PII masked

€1.2kspend

01 / gateway

Any model. One endpoint. Your rules.

Voban speaks the OpenAI API. Point your existing tools at Voban and switch from Mistral to Claude by changing one parameter. Same key, same Guard rules, same audit trail.

# any model, one endpoint
$ curl https://api.voban.ai/v1/chat/completions \
    -H "Authorization: Bearer sk-voban-..." \
    -d '{ "model": "mistral-large", "messages": [...] }'

# switch model. same key, same rules, same audit trail
$ curl https://api.voban.ai/v1/chat/completions \
    -H "Authorization: Bearer sk-voban-..." \
    -d '{ "model": "claude-sonnet", "messages": [...] }'

api

Drop-in OpenAI compatible

Works with the OpenAI SDK, Cursor, Continue and any tool your engineers already use. Change the base URL. That's it.

routing

Sovereign routing

Sensitive workloads automatically routed to EU-hosted models. Frontier models for everything else, with the same Guard policy.

budgets

Real-time budgets

Per-user, per-team, per-model spending tracked live. One gateway instead of five subscriptions. Onboard by email, revoke in one click.

02 / guard

Your AI policy, enforced on every request.

The Guard sits between your users and any model. It analyzes prompts, masks sensitive fields, enforces routing rules, and records every decision in an audit-grade log.

Inspect

The prompt is parsed before it leaves your infrastructure. PII, credentials, IBANs and prompt injection attempts are detected line by line.
Mask

Sensitive fields are masked or redacted according to policy. The original values never reach the model.
Route

The request is routed to the right model based on your rules: sovereign provider for regulated data, frontier for the rest.
Log

Every decision (what was masked, what was routed, what was blocked) is recorded with a tamper-evident timestamp and a stable request ID.
Return

The model's response is checked, logged, and returned to the user. Same answer. Complete audit trail.

without voban

Your legal team asks the AI to review a contract. The prompt contains client names, amounts, confidential clauses. The data reaches a US-hosted model. No trace. No audit trail.

with voban

The Guard detects sensitive fields, masks PII before the request leaves your infrastructure, routes to an EU-hosted model, and logs every decision, line by line, with a stable request ID your DPO can audit.

guard / recent events live

time	user	event	action
14:32:07	L. Martin	PII detected	3 fields masked → sent to Mistral Large
14:31:44	A. Chen	Injection blocked	Request rejected · admin notified
14:31:12	S. Duval	Rerouted to EU	Sensitive data → rerouted from GPT-4o to Mistral
14:30:58	J. Park	OK	Passed → Claude Sonnet

03 / sovereignty

Sovereign by design. Vendor-free by default.

The European answer to AI lock-in. Mistral, Qwen, Llama on EU infrastructure for your most regulated workloads. Frontier models (Claude, GPT, Gemini) wrapped in the same Guard protection, regardless of where they run.

Sovereign models

Mistral, Qwen, Llama on EU infrastructure. Data never leaves Europe. For your most regulated workloads: legal, HR, finance.

Frontier models, governed

Claude, GPT, Gemini behind the same Guard. PII masking, injection detection, full audit trail, regardless of where the model actually runs.

Cost control, line by line

Per-user, per-team, per-model budgets. One gateway instead of five subscriptions. Real-time spending visible to admins, not just to finance.

Zero vendor lock-in

Switch providers in one parameter. If a price changes or a service goes down, your teams keep working. No migration. No downtime. No re-tooling.

04 / blog

Thinking about enterprise AI.

Governance, sovereignty, architecture. What it takes to deploy AI without losing control.

April 2026

The Control Plane: An Architecture for Enterprise AI

What does the alternative to prohibition look like? Five layers, in the order a request meets them.

Read article

April 2026

Enterprise AI: The All-or-Nothing Trap

There are two ways to mismanage AI. Let it run wild, or lock it down. Both lead to the same place.

Read article

05 / get started

AI governance is no longer optional.

Bring your AI under control before your auditors ask. We onboard early teams hands-on, in days, not quarters.

Book a demo contact@voban.ai

// EU AI Act · GPAI transparency obligations apply since August 2025; high-risk system requirements take effect August 2026 (Regulation EU 2024/1689, Art. 113). Voban records every AI-assisted decision with a stable request ID, a tamper-evident timestamp, and a complete trace of every Guard decision.