Where an AI agent runs affects who can access its data, how it connects to enterprise systems, and who can stop it. Single-tenant deployment gives one organization a dedicated platform environment. It does not, by itself, guarantee that every connected service stays inside that environment.
We deploy Conduit in a client's cloud subscription or on premises. The purpose is to put agent operations within the client's access and operating model. Making that boundary real requires examining the complete data flow.
What belongs inside the deployment boundary
The platform stores operational state such as agent runs, approvals, knowledge, and audit records. Identity, secrets, permissions, and administrative access need explicit owners and controls.
Map more than the primary database. Backups, monitoring, support access, document processing, and model calls can create additional paths for information. A dedicated installation can still send a prompt to an external model provider if it is configured to do so.
Ask where each component runs, what information it receives, and what it retains. “Single-tenant” describes isolation; “on premises” describes location. Neither phrase answers every data residency question.
Choose inference deliberately
Some organizations can use approved external inference services. Others need models to run within their own environment. That choice affects infrastructure, available models, capacity planning, and ongoing maintenance.
Zeus is our on-premises AI infrastructure offering for select clients. We scope workloads and operating requirements before specifying a deployment. Model quality and performance still need to be evaluated on the work the organization intends to run.
What single-tenant operation costs
A dedicated deployment creates responsibilities that a shared service may centralize. Each installation needs a release process, backups, monitoring, incident response, and recovery testing. Upgrades must be checked against its integrations and representative workflows.
That work needs an owner whether the platform is operated by Current, the client, or both. Define support access, maintenance windows, service expectations, and escalation paths before production.
Current's operational access is designed to be revocable by the client. Revoking that access should be planned alongside the support and maintenance responsibilities the client will retain.
Questions to resolve before deployment
- Which systems may the agents read from and write to?
- Where do model inference, logs, and backups run?
- Who approves access and reviews administrative activity?
- How are releases tested and failed upgrades recovered?
- What data and configuration can the client export?
- What happens operationally when the engagement ends?
Record the answers in the architecture and service agreement. They are more useful than a general promise that the data never leaves a building.
For the controls applied to individual actions, read our guide to AI agent governance. To scope an installation, talk with Current.
