AI for CPA Firms: Document Requests and Filing Checklists

An AI agent can take over the repetitive middle of a CPA firm's document work: sending requests, tracking what came back, and checking each return against a filing checklist. Partners and managers still approve anything that gives client advice, changes a deadline, or leaves the firm. Set up this way, the agent also keeps the record a firm needs to show who asked for what, when, and what was done about it.

Key takeaways

• Document chasing is a strong first task for an agent because the rules are explicit: who owes what, by when, and what counts as received.
• The agent drafts and tracks requests, while staff approve any outbound message that touches judgment, fees or deadlines.
• A filing checklist is useful only when each line names an owner, a source document and a review step.
• A more capable agent needs permissions scoped per client, a log of every action, and a clear rule for when it must hand off to a person.
• Measure the work by open items cleared and exceptions escalated correctly, not by the number of messages sent.

Where document chasing costs partner time

Most firms track requests in some combination of email threads, portal uploads, spreadsheets and memory. The same reminder gets typed several times for the same client, a missing bank statement surfaces two weeks after it was due, and the partner learns about a gap only when the return is already stalled. None of this is a failure of effort. It happens because the work has no single list that everyone trusts, and because the follow-up depends on whoever remembers to do it.

An agent helps most when it works from that list instead of from memory. It can send the next reminder on the agreed cadence, log each receipt against the right request line, and tell the manager which items are still open and how long they have been open. The firm's own process decides what the list contains. The agent makes that process run without someone having to hold it in their head.

How to build the request list before anything is sent

The agent should not write its first request from a blank page. The firm defines the rules once, and the agent applies them to each client. A practical sequence looks like this:

• Pull the scope for each client from the engagement letter and the prior-year file, and note any services that were added or dropped.
• Write one request line per document, with the document name, the reason it is needed, the period it covers and the accepted format.
• Name the client-side owner for each line, such as the bookkeeper, the payroll contact or the owner personally.
• Mark each item as required to start work, required to finish, or conditional on something else in the file.
• Set the reminder cadence and the person who receives escalations when an item passes its agreed date.
• Have a manager approve the message templates and the cadence before the agent sends a single message.

This list becomes the agent's working memory for the engagement. Keep it in a system the firm controls, and treat changes to it as approvals, not as edits the agent makes on its own.

What a filing checklist needs to be useful

A filing checklist that says only "return reviewed" gives no protection and no information. Each line should be answerable with a yes or no and should point to the evidence behind the answer. For example, a line such as "Prior-year carryforward agrees to the source schedule" needs a reference to the schedule and the name of the person who checked it.

A checklist line is complete when someone else could verify it without asking the preparer.

The agent can track which lines have evidence attached, flag lines with no evidence, and remind the owner of each open line. It should not mark a judgment line complete on its own. Judgment items, such as whether a particular expense is deductible or whether a estimate is reasonable, stay with the reviewer who signs off on the return.

Illustrative scenario: a client's bank statements arrive for eleven months of a twelve-month period, and the missing month is the one with the largest unusual transaction. The agent logs the gap, reminds the client contact once on the agreed schedule, and escalates to the manager with the missing item named and the request history attached. The manager decides whether to call the client, and the agent does not write anything that interprets the transaction.

Where the agent stops and a partner decides

The boundary between agent work and partner work should be written down before the agent is switched on. A simple decision guide helps:

• The agent may send approved standard reminders, log receipts, index uploaded files to the correct client folder, and update the status of request lines.
• The agent must route to a person any message that interprets a tax position, answers a question about a notice from a tax authority, changes a deadline, or proposes a new engagement or fee.
• The agent must never file a return, submit a payment, change a figure in a prior-year file, or send a client a document it has not been approved to send.

If a client replies with a question that is not covered by the approved templates, the agent should acknowledge receipt in a neutral way and hand the thread to a named person with a short summary. Speed matters, but a fast wrong answer on a tax question costs more than a same-day handoff.

What a more capable agent would need in permissions and audit records

A first version of this work can run with read access to the document folders and write access only to the request list and the messaging channel. As the agent takes on more, its permissions should grow in small steps and each step should be reviewed. Access should be scoped to the clients assigned to a given engagement team, not to the whole firm. Delegated access should follow the principle of least privilege, the same idea that OWASP describes in its guidance on authorization: every action is checked against what the specific user or agent is allowed to do on that specific record.

Client-supplied documents need a particular kind of care. A scanned letter or a PDF can contain text written to influence an automated reader, whether by accident or on purpose. The OWASP guidance on prompt injection prevention is clear that content an agent reads should be treated as data, not as instructions. In practice, this means the agent never acts on a sentence it finds inside a client document, and it reports unusual content to the reviewer.

The audit record is what makes this defensible. Each action should log who or what acted, which client and request line were affected, the time, the source document or message, the template used, and any approval given. Repeated actions need protection too. If a retry sends the same reminder twice, the client sees a duplicate and the record becomes unreliable. Stripe's documentation on idempotent requests shows one provider's approach to making a repeated instruction produce a single result, and the same idea applies to reminders and status updates. Good traces of agent activity follow the general pattern in OpenTelemetry's observability concepts, where logs, traces and metrics let a team reconstruct what happened and why.

How Be Human runs this as a managed service

Be Human designs these agents around one defined role, such as a document coordinator for a specific engagement team. The first engagement defines the request rules and checklist structure, the memory the agent needs, the tools it may use, the approval boundaries and an observable success criterion, such as the share of request lines closed on the agreed schedule. Managed operations then cover monitoring of the agent's activity, recurring evaluations against the firm's own examples, permission reviews as access changes, integration maintenance with the firm's existing systems, and incident response if something goes wrong. The service is delivered remotely, in English and Portuguese, and the firm keeps control of its identity, memory and channels throughout.

How to measure whether the agent is helping

Choose two or three measures before the agent starts. Useful ones include the number of request lines closed by the agreed date, the number of exceptions escalated to a person, and the number of manual corrections a reviewer had to make to the log. Review these with the manager every few weeks. If the exception count rises, the cause is usually an unclear request line or a template that does not match how clients actually reply, and both are fixable.

If you want to find the first task in your firm that is worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent send document requests to clients without partner review? It can, provided the firm has approved the message templates, the reminder cadence and the escalation rules in advance. Any message that goes beyond those approved texts, such as a reply that interprets a tax position or changes a deadline, should be routed to a named person before it is sent. Reviewing the templates once, and sampling the agent's sent messages on a regular schedule, keeps this arrangement sound.

Should an AI agent file returns or submit anything to a tax authority? No, not as part of a document coordination role. Filing and payment are high-consequence actions that depend on professional judgment, signatures and the firm's own controls, so they belong outside the agent's permissions. The agent can prepare the checklist evidence that a preparer or reviewer relies on, but the submission itself stays with a person.

What client information should stay out of an agent's memory? Keep out anything the firm would not want a model to retain or reuse, including full identification numbers, account credentials and unrelated client files. The agent needs the request list, the client's reporting period, the accepted formats and the contact roles. Store sensitive values in the firm's own systems and let the agent reference them only when a specific task requires it, with access logged.

How is an agent different from the practice management software the firm already uses? Practice management software records work and provides a central file, but it does not usually act across email, portal uploads and reminders in a consistent way. An agent can read those channels, update the request list, draft approved messages and escalate exceptions, while the software remains the system of record. The two work best together, with the agent operating inside boundaries the firm has written down and reviewed.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.