Evaluate self-hosted AI with clear deployment responsibilities, workload budgets, model and tool policies, audit evidence and operational costs.
Self-hosting changes who operates the product, not whether governance is needed. Use this guide to decide between managed Sluis and Edge, then assign ownership for policy, spending and operations.
Choose the deployment boundary you actually need
Sluis offers a managed EU-first gateway and Workspace. Edge runs the full product on your own Kubernetes or Docker environment, including air-gapped deployment with local models.
Illustrative scenario: a research team cannot send restricted material outside its network. Evaluate Edge with approved local models, but verify capacity and model suitability rather than assuming managed-provider capabilities remain available offline.
- Document whether external processing is permitted.
- Select models suitable for the intended deployment.
- Test required workflows without external network access.
Budget for ownership, not just installation
With Edge, the customer operates the environment, including backups and licence management. A deployment package does not remove responsibility for infrastructure, recovery planning or maintaining an operable installation.
Assign ownership for updates, monitoring and incident response in your implementation plan. Confirm support boundaries and licence arrangements rather than assuming a self-hosted product is a managed service or a hardware appliance.
- Name owners for infrastructure, backups and licensing.
- Run a recovery exercise before sensitive production use.
- Confirm support boundaries and offline operating requirements.
Control models, tools and workload spending together
Sluis supports model and MCP tool policy, allow/deny modes, budgets per workload and rate limits. Define policies around approved work rather than relying on a company-wide model list alone.
A workload budget and a rate limit answer different questions: spending allocation and request pace. Test their behaviour with realistic traffic and confirm what happens at each configured boundary.
- Assign each workload an owner and budget.
- Specify permitted models and MCP tools explicitly.
- Test budget and rate-limit boundary behaviour.
Use audit evidence without overstating observability
Sluis provides a hash-chained audit trail. This supports review of recorded events, but does not by itself establish complete application tracing, automatic incident resolution or visibility into unapproved tools outside Sluis.
Define the evidence needed for spending reviews and policy investigations. Check recorded fields, access and retention against that brief; consult current usage-based pricing for managed services and confirm Edge licensing separately.
- Test an audit review against a representative incident.
- Verify evidence access and retention responsibilities.
- Review workload spending and policy exceptions regularly.
Questions before you decide
Does self-hosting make AI sovereign?
It can change control and data flows, but sovereignty also depends on ownership, dependencies, models and legal exposure. Use the existing sovereign AI due-diligence guide for deeper assessment.
Can Edge operate without internet access?
Edge supports air-gapped operation with local models. Validate required workflows, capacity, maintenance and licence arrangements for your environment; do not assume external model services are available offline.
Do governance controls eliminate shadow AI?
No. Controls apply within the configured Sluis environment. Address unapproved tools through organisational policy, employee guidance and appropriate external controls, without assuming universal discovery.