A mid-tier audit and assurance practice
The firm wanted AI in the audit, but client data could not touch a third-party cloud. So we built a desktop application with the models running locally. The intelligence sits on the auditor’s machine, and the data never leaves the firm’s perimeter.
Representative engagement. Numbers are directional outcomes typical of builds of this type, not figures from a single named client account.
Audit work is repetitive, high-stakes, and bound by confidentiality. Juniors spent days sampling ledgers, ticking and tying, and drafting working papers by hand. The partners knew AI could lift the manual load, but uploading a client’s general ledger to an external API was a non-starter for engagement letters and regulators alike. Every off-the-shelf tool assumed the cloud.
We inverted the usual architecture. Instead of sending data to the model, we shipped the model to the data. A native desktop application runs the AI modules on the auditor’s own machine, with no client financial data leaving the device. Anomaly detection, sampling, and working-paper drafting all happen locally, and the firm keeps a full audit trail of what the AI suggested and what the human accepted.
Juniors now start from a flagged, sampled, partially drafted file instead of a raw ledger. The firm gets the speed of AI with none of the data-residency risk, because nothing leaves the building. Partners review the same way they always have, just over a much shorter pile, and the firm owns the application outright.
The first question every partner asked was where the data goes. The answer is nowhere. That is why we could actually use it.
Partner, audit and assurance firm
