Private AI can mean several different things. A vendor may be describing a dedicated environment, a private network route, a model running on customer-controlled infrastructure or a service with particular data-handling terms. Those details determine whether the proposal fits your business.
For a construction supplier, the workflow may involve a plant computer, a cloud accounting system, delivery documents and payment records. Ask about the whole path, not only the part labeled private.
Where is source information stored
List the applications and records involved. Identify what Nexus or another service reads, copies or retains. Include the documents used for retrieval, account identifiers and any derived data. A diagram is useful when it shows the actual deployment rather than an idealized product picture.
Where does model inference happen
Inference is where the model processes the request. Establish the service, account and region that perform that work. A model reached through a private endpoint may still run in a provider-managed environment. Private connectivity describes a network path; it does not by itself establish ownership of the model host.
Who can access the workflow
Consider employees, administrators, service accounts and support staff. Ask how location-level access is enforced and how access is removed when responsibilities change. An agent should not expose records a user is unable to access through the approved workflow.
The right answer is specific enough for your IT team to verify. Labels such as enterprise grade or secure by design cannot replace an explanation of identities, permissions and responsibility.
What is retained after a request
Review prompts, outputs, retrieved documents, operational logs and backups. Ask who retains each item, for how long and for which purpose. Model-provider terms can vary by service and configuration; a private route should not be interpreted as a universal promise of no retention or no outside processing.
Which integrations cross the boundary
An agent may call external software, fetch a document or send an approved update. Identify those connections and what data they carry. Include authentication services and monitoring systems. A tightly controlled model environment can still participate in a workflow that communicates with other providers.
Who maintains and recovers the environment
Agree who handles updates, monitoring, incidents and recovery. Ask what happens if the local connection fails or a model service becomes unavailable. Test the recovery steps that are important to the selected workflow, including how incomplete transactions are identified.
What exactly has been validated
Request evidence for the offered deployment and services. A certification for one payment service does not automatically describe every data or AI component. Similarly, a successful demonstration does not establish that all network and access controls have been tested in production.
Nexus offers private deployment options that should be scoped with the customer’s IT requirements. Nexus Edge connects supported on-site software to the platform; it is not evidence that model inference runs on that same machine. Privately hosted models, external model services and customer-controlled environments need their own clearly documented scope.
Before choosing a private AI deployment, ask for a data-flow discussion covering these seven questions. The result should explain where information goes, who can act on it and who is responsible when something fails.