What does a modern AI stack look like for a small CPA firm doing bookkeeping and tax work, and how can client data privacy be maintained?

One firm owner (a former software engineer turned CPA running a ~10-person small business accounting/tax/payroll firm in Texas) described their current AI stack: Claude as the daily-driver LLM, Perplexity for tax law research, and a self-hosted 'OpenClaw' setup for client communications. For data privacy, the firm runs two local DGX Spark units hosting local LLMs in-office so client data never leaves the premises, keeping terabytes of client information local. This local setup is used to help support SOC 2 compliance. The firm is also building its own tool, referred to as 'XBub AI,' intended to eventually replace QuickBooks for client cleanup work. As an example workflow, they showed how a scanned IRS notice can be run through their AI to extract key fields automatically into a database, so the team knows who should respond and by when. The presenter noted that larger context windows (up to a million tokens) in newer models have been a major recent shift, allowing ingestion of a client's entire document history at once (PDFs, images, a full year of data) rather than feeding documents one at a time, which has improved the reliability of their automations over the last six months.