How a CPA practice runs client document collection with AI agents.

Every accounting practice runs the same hidden second job: chasing clients for documents. The shoebox of receipts, the missing March bank statement, the invoice list that lives in a spreadsheet only one person understands. Industry blogs love to speculate about how AI will fix this by 2030. This article is not a prediction. It is the blueprint of a system we built and operate today, agent by agent.

The shape of the system

A two-sided workspace connects the practice and its clients. The client gets a dead-simple room: drag a file in, done. The practice gets the organized result: documents filed by month, a checklist that knows what is missing, and findings that surface money problems before they compound. Between those two rooms sits a small fleet of AI agents, each with one job and hard limits.

The agents, by role

The Reader. When a client drops a receipt, an invoice, or a messy spreadsheet, the extraction agent reads it into structured records: vendor, amount in exact cents, date, category. The client's workbook never needs to be reformatted; the agent meets the mess where it lives. A confirmation step keeps a human in the loop before any schema is trusted.

The Money Watcher. A rules agent reads the promoted records continuously and raises findings a bookkeeper would want flagged: an invoice past due long enough to matter, a duplicate charge, a total that does not foot. Every finding is confirmable or dismissible by a person, and a dismissal is permanent; the agent never nags twice about the same judgment call.

The Requester. When the practice needs a document the client has not sent, a request agent places it in the client's room, emails them, and blocks the month from being marked ready until it is resolved. The chase, automated, without a single “just following up!” email written by a human.

The Lifecycle Nudger. Sign-ups that stall, clients who never upload, trials about to lapse: a lifecycle agent sends exactly one well-timed note per situation, once ever. Persistence without spam is a policy decision, and it is enforced in code.

The Auditor. Every model call the system makes is written to a ledger: which prompt version ran, what it cost, how long it took, whether it failed. Never the document contents. When a client asks “what does your AI do with my data,” the answer is a screen, not a shrug. This is the part most agent blueprints skip, and it is the part a regulated profession actually requires.

The human gates

Agents do the work that scales: reading, filing, requesting, nudging, flagging. People keep the three decisions that matter: confirming a schema before it is trusted, judging whether a money finding is real, and closing the month. The design principle is the same one we apply to every agent system we build: separate the infinitely scalable work from the small set of judgment calls a professional should always own.

What it took to build honestly

The agent logic is the small part. The majority of the engineering went into what surrounds it: tenant isolation proven by adversarial tests on every deploy, an audit trail on every action, browser journey tests that click what a real bookkeeper clicks, and email that confirms instead of vanishing. A blueprint without that surrounding plane is a demo. With it, it is a product a CPA can put in front of clients.

This system runs today at SheetTamer, free for practices. And the pattern generalizes: the same agents-plus-gates architecture runs our paid search, our customer service desk, and our clinical AI governance work.