Accounting Firm AI Document Assistant

The situation

The firm employed between 10 and 40 people across bookkeeping, tax prep, and client advisory. Their working knowledge lived in scattered places: a shared drive of standard operating procedures, engagement templates, an internal policy handbook, and years of prior filings and workpapers. Finding the right answer often meant interrupting a senior staff member or digging through folders during the busiest weeks of the year.

Two problems compounded each other. Routine questions ate up billable time, and some staff had started pasting client figures, names, and document excerpts into public AI tools to draft responses or summarize procedures. That put sensitive financial data outside the firm's control and created exposure the partners could not see or measure.

What we looked at

We started with a short discovery: which documents people actually reach for, where those documents live, and which questions come up again and again. We catalogued the approved sources (SOPs, the policy handbook, engagement letters and templates, and selected prior-filing references) and flagged material that should never be exposed, such as raw client PII and anything under active dispute.

We also reviewed how staff were already using AI in the wild. That told us the assistant had to be at least as fast and convenient as the public tools it would replace, or people would quietly keep using those instead.

What we built

We built a private AI knowledge assistant that answers questions only from the firm's approved documents. A document ingestion pipeline pulls in the curated SOPs, policies, and templates, splits them into passages, and indexes them for retrieval. When a staff member asks a question, the assistant uses retrieval-augmented generation (RAG) to pull the most relevant passages and answer from them, with citations back to the source document and section.

Access is gated by role. The model runs against the firm's data inside controlled boundaries rather than a public endpoint, and the index excludes the sensitive material we flagged during discovery. Answers that fall outside the approved sources return a clear 'not found here' rather than a confident guess.

Alongside the tool, we wrote a plain-language AI governance policy covering approved tools, what data may and may not be entered, and how to handle client information, then delivered role-specific training so each team knew the rules and the workflow that fit their job.

How it works

A staff member asks a question in plain language, such as how to handle a particular adjusting entry or which template applies to a new engagement type. The assistant retrieves the matching passages from the approved index, drafts an answer grounded in those passages, and shows the citations so the answer can be verified against the source.

Because retrieval is scoped to approved documents and access follows role, people get a sanctioned place to ask routine questions without reaching for an outside tool. When source documents change, re-ingesting them keeps the assistant's answers current.

Results

Staff got answers to routine policy and procedure questions in seconds rather than spending part of an hour searching or waiting on a colleague. Just as important, the firm now had a sanctioned alternative to public chatbots, which sharply reduced the temptation to paste client data into tools outside its control.

These figures are illustrative and estimated for this scenario rather than audited, but the direction was clear: less time lost to searching, and a much smaller surface area for sensitive data to leak.

Why it matters

For a firm handling client financials, the risk is not just slow answers, it is where the data goes when people improvise. Grounding the assistant in approved sources, scoping access by role, and backing it with a clear policy keeps client information inside approved boundaries while still speeding up everyday work.

This is governance guidance and a set of review workflows, not a certification or a guarantee of compliance. It gives the partners visibility and a defensible default, and gives staff a faster, safer way to do their jobs.