Compare cloud, private and local AI by tracing prompts, retrieved text, logs and support access. Deployment labels alone do not establish data control.
Before choosing cloud, private or local AI, trace the information required for one task. Labels such as “sovereign” or “hybrid” do not describe every processor, stored copy or person who can access that information. A useful comparison records the actual flow and the terms that apply to it.
A local document store can still feed a cloud prompt
Consider an illustrative professional-services workflow that drafts a summary of a client contract. The documents and search index stay on an office server. When a user asks a question, the application retrieves contract passages and sends them to an external model to compose the answer.
That design keeps the source repository local, but selected contract text still leaves it. Calling the architecture hybrid does not change that fact. If the requirement is that contract content must not be sent to that provider, the proposed workflow has not met it.
Map the individual stages before deciding whether the arrangement is acceptable:
| Stage in this example | Data to account for | Evidence to request |
|---|---|---|
| Import and search indexing | Original text and derived index data | Where processing occurs and who can query it |
| Answer generation | Question and retrieved contract passages | Exact destination service, terms and settings |
| Application diagnostics | Requests, errors and possibly response text | What is logged, who sees it and retention rules |
| User access | Drafts and saved conversations | Identity checks, sharing and deletion behaviour |
| Support and recovery | Backups, support exports and troubleshooting access | Operator permissions and applicable locations |
The table is an assessment aid, not a description of all products. Test it against a specific configuration and fill unknown entries before introducing the relevant data.
Separate training from other data handling
The OAIC’s guidance on commercially available AI products notes that some products have terms or settings allowing inputs to be used for further training. Check the exact product and feature rather than assume all providers use inputs in the same way.
A commitment not to train on inputs does not, on its own, answer whether requests are retained, reviewed for abuse, copied into support records or processed by another provider. Ask separate questions about each operation and record the contractual answer alongside the configuration that enforces it.
For organisations covered by the Australian Privacy Principles, the same OAIC guidance addresses obligations and good practice when AI uses personal information. Determine which requirements apply to the entity and use case. A hosting region or a vendor’s reputation is not a complete privacy assessment.
Compare complete operating arrangements
A hosted service may reduce the infrastructure your team operates, while introducing contractual and technical dependencies on the provider. A local model gives you responsibility for more of the operating environment, including capacity, updates, access controls, backups and recovery.
Neither arrangement automatically preserves confidentiality. In the contract example, a local model connected to an external tracing service may send text through its logs. A cloud workflow with a verified narrow data flow may be suitable for an approved task, subject to its requirements and agreements.
Test each candidate on representative authorised material. Include extraction quality, response quality, total cost and staff effort. A deployment that meets the location requirement but cannot produce usable results has not met the whole task. Likewise, a capable answer does not excuse a prohibited transfer.
Verify the boundary and the exit
Use fabricated documents with recognisable markers to exercise the workflow. Inspect configured destinations and available request, storage and logging evidence. Disable unneeded integrations and confirm what happens when an external dependency is unavailable. Document any visibility limits that prevent a complete verification.
Treat a change of model, feature or support process as a reason to review the flow again. Keeping business data in a separate repository can help portability, but prompts, indexes, permissions and integrations may still require migration work.
Before committing, establish how to export required records, revoke access and handle remaining copies under the applicable retention terms. Test an export with a representative record; a statement that the data is yours does not prove it can be retrieved in a usable form.
Bring the task and the completed flow table to a sovereign AI assessment or integration review. Specify the control you need in observable terms, then evaluate the design against it.
