Assess private AI against client information barriers and the full data path
Assess private AI against client information barriers and the full data path.
QuantalAI helps professional-services firms assess private AI for a defined preparation or research task. Self-hosting can provide control over a model deployment, but confidentiality also depends on retrieval, document processing, logs, backups, support access and external tools. The assessment starts with the firm’s client commitments and information barriers, then tests whether the proposed system can meet them while producing useful work.
Book a discovery callWhat the implementation covers
Complete data path
Identify where files, prompts, embeddings, outputs and operational records are processed and stored.
Engagement boundaries
Apply current user permissions to retrieval and related outputs, including restricted prior engagements.
Task evaluation
Test evidence support, omissions and review effort using material approved for the assessment.
Operating model
Compare infrastructure, licences, maintenance, availability and staff effort across suitable options.
Define what private must mean for the task
Start with the information and commitments involved. A firm may require local processing, specified operator access, restrictions on reuse, engagement-level permissions or a combination. Translate those requirements into conditions that can be checked.
The OAIC’s AI privacy guidance calls for due diligence on intended uses and access to personal information. A deployment label does not replace that assessment or resolve the firm’s separate client confidentiality commitments.
An illustrative engagement-research assistant
Suppose consultants want to find an approved framework and relevant prior work. The system should distinguish reusable firm material from client-specific evidence that the user cannot access or repurpose. This is a proposed design example, not a claim about a deployed service.
Test a consultant assigned to engagement A but excluded from engagement B. Include similar project names and a shared framework referenced in both. The assistant should retrieve the permitted framework without exposing B’s client facts, excerpts or document title where that title is restricted.
Test a revoked assignment and saved answers too. Removing access to a source may not automatically remove copies already placed in conversation history, exports or caches.
Map every service that handles the material
| Component | Question to resolve |
|---|---|
| Document extraction and embeddings | Where do the document contents go before retrieval? |
| Model and tools | Which services receive prompts, excerpts or action inputs? |
| Logs and support | Who can access records, and what do they contain? |
| Backups and saved outputs | Where are copies retained, restored or exported? |
Inspect actual settings and test network behaviour where appropriate. Hosting the language model locally does not establish that the surrounding application is local. ASD’s deployment guidance addresses security across deployment and operation.
Evaluate the work the model produces
Use material approved for testing and a set of tasks that includes missing, conflicting and outdated evidence. Check whether drafts preserve qualifications, cite supporting sources and expose uncertainty. Measure the time a qualified reviewer needs to correct them.
Grounding a response in internal files does not guarantee accuracy. A model can omit a condition or combine facts from different contexts. Accept the system only against requirements established for the particular preparation task.
Compare the operating choices
Assess model licence terms, hardware or cloud capacity, concurrency, patching, monitoring and recovery. Self-hosting may suit a defined need, but its total cost includes people and availability requirements. A managed deployment may suit another task if its terms and controls meet the firm’s requirements.
See sovereign AI for the broader assessment and open-weight models for model considerations. Agree ownership of access reviews and future changes before expanding a pilot.
Read more about our Sovereign AI service and our work in Professional Services sector.
Illustrative solution designs.
Frequently asked.
Does self-hosting guarantee that client material stays inside the firm?
Can every consultant search all prior engagements?
Is a small open-weight model always sufficient and cheaper?
Does using an Australian cloud region settle all privacy requirements?
Assess one workflow before committing to the build
Bring the intended task, source systems, users and constraints. We can define the implementation, acceptance evidence and responsibilities for operating it.
Book a discovery call