Compare four different approaches
One approach is to exclude sensitive information from the task. Redaction, approved reference material, or nonsensitive examples may be enough. Another is an enterprise cloud service with contractual protections and administrative controls. A third is a dedicated hosted environment. A fourth is running a model on equipment your organization controls.
These approaches have different costs, capabilities, maintenance burdens, and data flows. A business cloud offering may process information externally even when it does not train models on your business data by default. OpenAI and Microsoft describe relevant protections for their eligible work offerings, but those statements must be applied to the exact plan and service. [S15, S16]
Trace the entire path, not just the model
Map where prompts, retrieved documents, outputs, logs, backups, and support records go. Include integrations, search providers, analytics, telemetry, and external tools. A locally running model can still be part of a workflow that sends information to a cloud connector.
Ask who has administrator access, how data is encrypted, what is retained, and how deletion is handled. Review subcontractors, geographic requirements, and the actual contract where those issues matter. Do not treat the word "private" on a sales page as an answer to these questions.
Understand what local hosting adds
Local hosting may help when keeping information within a defined environment is an important requirement. It also creates work: model selection, hardware capacity, patching, access control, logging, monitoring, backups, and incident response. The business needs someone responsible for those tasks.
A smaller local model may perform differently from a hosted service. Test the work you intend to use it for and the consequences of a bad answer. Hosting location does not fix fabricated content, excessive permissions, or an unsafe approval process.
Reduce access before increasing complexity
Start with the smallest approved data set and read-only access where feasible. If a workflow can send messages or update records, require explicit authorization for those actions and evaluate the permissions separately from document access. OWASP identifies excessive agency as a risk when AI systems receive unnecessary functions, permissions, or autonomy. [S14]
Choose architecture from the requirement
Identify whether the actual need is confidentiality, restricted retention, data residency, isolation, offline operation, or something else. Those are different requirements and may lead to different solutions. Blue Tree Technology® can discuss your use case and help define what must be verified before implementation. Do not buy a more complex platform until the business can explain which risk or operating need it resolves.
Verification sources
S15 OpenAI: How data is used to improve model performance
S16 Microsoft: Enterprise data protection in Copilot and Copilot Chat