Private AI for internal audit workpapers
AI saves real time on the repetitive part of internal audit documentation, and a self-hosted model lets you use it without sending findings to an AI vendor. That matters because a workpaper is one of the most candid documents a company produces about itself: a control that isn't operating as designed, a sample that didn't reconcile, a gap nobody has fixed yet. Most companies don't want that on a vendor's server before management has remediated it.
What a workpaper documents
A workpaper records the control being tested, the sample selected, the evidence reviewed, and the auditor's conclusion, including any exception found. Across a full audit cycle, the set is a detailed map of where the company's financial controls are weakest: a segregation-of-duties gap, an access review that slipped its schedule, a reconciliation still done by hand because the system integration was never finished. None of that should circulate before the finding is remediated and closed.
Pasting excerpts of that material into a general-purpose AI tool to help write up a finding sends unremediated control weaknesses to a third party's infrastructure, under a retention policy written for general use, not for audit evidence. That's a different category of exposure than most of what goes through a shared AI tool day to day.
What changes when the drafting model is self-hosted
Running the drafting assistant on a dedicated DGX Spark means workpaper drafts, finding language and sample documentation go to a model the audit function runs itself, on a machine assigned only to it. Nothing goes to an AI vendor's API, so no vendor retention or training policy applies. The auditor still gets help turning testing notes into a formatted workpaper, or a finding into clear remediation language for management.
A rented Spark is still GPUwerk's hardware in GPUwerk's rack, operated under its published DPA. Check that your audit methodology and any client confidentiality terms allow a hosting provider before real audit evidence goes on it.
A practical setup is Open WebUI configured with the audit function's workpaper template and prior-year findings as reference, restricted to the audit team rather than the broader finance organization. If the team keeps a library of control narratives and past findings (the approach in the guide to on-premise RAG), the model can check a new finding against how a similar issue was written up before, so severity language stays consistent.
What the model is useful for, and what stays with the auditor
A drafting model helps format workpapers to a consistent standard, summarize a large sample of test results into a narrative, and draft first-pass remediation language. It does not replace the testing, or the judgment about whether a control gap is a reportable finding. That judgment and its evidence have to trace back to work the auditor performed, not to a model with no access to the systems under review.
What this doesn't solve
Self-hosting the drafting model doesn't satisfy independence or documentation requirements under whatever audit standards apply, doesn't replace audit committee reporting, and doesn't decide who inside the company should see a finding before remediation is complete. Those remain audit-function policy questions. What it removes is one specific exposure: a record of unremediated control weaknesses sitting with an AI vendor, under a retention policy written for general use, before management has had the chance to fix them. See pricing for what a dedicated Spark costs.