On-premise
Deploy the agent on your infrastructure.
When the application itself stays with you, the agent can be deployed there rather than consumed as an external service. This option requires an assessment of your prerequisites.
Two different things, often conflated
Hosting the application on your side and running inference on your side are two separate decisions. The agent can be deployed on your infrastructure while calling an external model, or the reverse.
Conflating them leads to disappointed expectations: a local deployment does not by itself guarantee that no data leaves, if the model being called is remote.
The assessment therefore starts by clarifying which of the two goals you are pursuing — often both, but not under the same constraints.
What needs examining
Technical prerequisites
Runtime platform, storage, network, and compute capacity if inference is local. These needs are quantified before any commitment.
Split of responsibilities
Who operates, who supervises, who intervenes on incidents. Deploying on your side moves part of the operational load.
Updates
How new versions arrive, at what cadence, and what happens if a version is not applied.
Disconnected operation
Some functions — external monitoring, for instance — assume outbound access. What remains possible without it is defined function by function.
This option is assessed case by case. We publish no reference configuration and no typical timeline: they depend entirely on your infrastructure.
Frequently asked questions
Do we need GPUs?
Only if inference runs on your side. A local deployment of the agent calling a managed model does not require them. Sizing then depends on the model and the volume.
Who provides support?
That is a contractual point, not a given. Deploying on your side means defining intervention levels and the access needed for diagnosis.
Can we start differently?
Often, yes. Validating usage on managed cloud and then assessing installation lets you know what you are deploying before deploying it.