An AI system can use a downloadable model, run on a rented server and call an external service for part of its processing. Calling it “private” without explaining that journey tells you little. Choosing an architecture requires examining the complete system.
Distinguish components and their licences
The licence of software that runs a model is not necessarily the model’s licence. Weights, data, extensions and a hosted offering may be subject to different terms. Before commercial use, read the licences for the versions actually selected.
The Open Source Initiative defines open source AI around freedoms to use, study, modify and share, together with the elements required to exercise those freedoms. The ability to download weights alone does not describe every right or all the information available. We favour a precise inventory over a label applied to the entire system.
Map the data flows
Follow a request from the browser to the result. Where are files stored? Which service converts documents into text? Where is the index produced? Which model receives the passages? Which logs retain the content, and for how long?
This diagram reveals dependencies that are sometimes overlooked: optical character recognition, translation, monitoring, backup or authentication. Running the model locally does not ensure that all those services remain in the same environment. The contract and configuration should match the journey actually deployed.
Size the infrastructure around usage
The hardware required depends on the model, concurrent requests, document length and acceptable response time. A demonstration for one person does not measure the experience of several teams at peak usage.
Before investing in a server, test a representative workload. Measure memory use, waiting time and behaviour when the service is saturated. Include updates, storage, backups and monitoring in the cost. Owning equipment still requires operating it.
Enforce access within the application
The model is not an authority for access decisions. The application must verify identity, roles and document permissions before transmitting data or executing an action. ANSSI’s guide addresses the security of a generative AI architecture as a whole, beyond model choice alone.
For a project, we suggest testing the boundaries: two users from different departments, a revoked account, a deleted source and an unavailable connector. These scenarios provide more useful evidence than a message simply asserting that data is protected.
Prepare an exit from the outset
Portability must be tested in practice. Can you export the documents, configurations and data you need to retain? Are the formats readable without the original provider? Who holds the operational credentials? How can unnecessary copies be removed?
Good documentation lets another team understand the components and routine operations. Open choices can support that independence, but they do not replace skills or a recovery plan. Our ecosystem presents tools to assess against these criteria, while our AI infrastructure pricing guidance distinguishes the initial project from operating costs.
