AI Agents for a Small CPA Firm: Client Requests and Deadlines

An AI agent can safely take over the routine layer of a small CPA firm's client communication: logging incoming requests, chasing missing documents against a checklist, and drafting deadline reminders that a staff member approves before they go out. It should not prepare tax positions, sign or submit filings, or answer questions that depend on a licensed professional's judgment about a specific client. The practical line is simple: the agent handles work that has a clear owner, a checkable outcome, and a boundary that staff can enforce.

Key takeaways

• Start with request intake and reminders. They are repetitive, text-heavy, and easy to check against the file.
• Let the agent draft, log, track, and organize, but never submit, sign, or advise on a tax position.
• Every deadline needs one named owner, one source calendar, and a person who approves reminders before clients receive them.
• Filings, payment authorizations, notices from tax authorities, and engagement changes stay with licensed staff.
• Judge the agent on a narrow measure, such as whether each reminder matches the open items in the file at send time, and review that measure on a fixed schedule.

Why requests and deadlines are the right starting point

Most small firms lose time in the same place. A client asks a question by email, sends a photo by text, uploads a file to the portal, and the missing bank statement sits unnoticed until the deadline is close. This work is repetitive and text-heavy, and its main risks are misrouting and silence, not a wrong tax answer. Both risks can be designed out with a clear role, which makes this a better first role than anything that touches the return itself.

Requests and reminders also create a record. Each message, chase, and reminder becomes an entry against the client file, which helps the preparer and the reviewer. An agent with a defined role keeps that record current, so nobody has to rebuild it from memory at the end of a busy week. Firms with intake-heavy work often see the same pattern, as described in AI Agents for Law Firms: Setup Guide for Intake and Deadlines.

Which client requests an agent can handle now

These tasks suit an agent with read access to the client file and write access only to the communication log and the draft queue:

• Acknowledge incoming client messages and log each one against the correct client and engagement.
• Ask for missing documents using the checklist that belongs to that engagement type, naming only the items still open.
• Draft deadline reminders from the firm calendar and place them in a queue for staff approval.
• Answer status questions only from fields a staff member has marked as current, such as whether a return is with the preparer or with the reviewer.
• Route any question about tax treatment to the assigned preparer with the full thread attached and a short summary written out.
• Prepare a weekly summary of what was received, what is still outstanding, and who owns the next step.

The agent should not estimate a refund, predict an outcome, or explain why a figure changed. If the record does not contain a current answer, the correct response is that the question has been passed to the preparer, and the agent should say so plainly.

Which filings and judgments stay with licensed staff

Some work belongs to people who hold licenses, sign documents, and carry professional responsibility. A useful test is to ask whether the output would be relied on as the firm's professional judgment or as a legal submission. If the answer is yes, the agent may prepare material, but a licensed person decides and acts.

• Preparing or changing tax positions, elections, or classifications for a client.
• Signing, submitting, or authorizing any return, extension, payment, or amended filing.
• Responding to notices or correspondence from tax authorities.
• Deciding whether a late filing, penalty, or reasonable cause argument applies.
• Confirming that a person who claims to act for a client has the authority to do so.
• Changing engagement scope, discussing fees, or agreeing to exceptions to the firm's terms.

Where the boundary is unclear, the default is that staff decides. The agent can still do useful work in that case by sending the thread to the right person with the question written clearly, which saves reviewer time without making the call.

A reminder workflow for one deadline, step by step

Illustrative scenario: a four-person firm has a recurring document deadline each season for a group of business clients. The owner wants reminders to go out without someone retyping the list each time, but does not want the agent making judgments about the documents. The workflow below keeps staff in the approval seat.

• Step 1: Staff confirms the deadline, the client list, and the document checklist for each engagement type in the calendar the agent reads.
• Step 2: The agent builds one reminder per client that names only the documents still missing from that client's file.
• Step 3: The drafts enter an approval queue. A staff member edits or approves each batch before anything leaves the firm.
• Step 4: After sending, the agent logs the time, the recipient, and the open items, and creates a follow-up task for the assigned staff member.
• Step 5: If a client replies with a question about treatment, timing, or an exception, the agent stops, attaches the thread, and assigns it to the preparer.
• Step 6: Each week, staff reviews the open items list, closes what has been resolved, and corrects anything wrong so the next reminder is accurate.

The approval step is short when the drafts are specific. The client receives one accurate message instead of a generic notice, and the preparer sees the gaps while there is still time to close them.

How permissions and the audit trail keep the agent in bounds

Client tax data is among the most sensitive material a firm holds, so the agent should operate with the smallest set of permissions that still does its job. That means read access to the relevant client folders, write access to the communication log and draft queue, and no access to filing software, payment systems, or signature tools. Each action should be checked against a defined role rather than inherited from whoever configured the system. The OWASP authorization guidance describes this principle of least privilege well, and it applies directly here.

Incoming content deserves its own caution. A client email, an uploaded PDF, or a pasted message may contain text written to steer the agent, such as a request to move a deadline or send a file to another address. This is the kind of risk described under prompt injection. Treat that content as material to summarize, never as an instruction. Log every action with its source message, the draft produced, and the person who approved it, so a reviewer can reconstruct what happened. The NIST AI Risk Management Framework is a practical reference for naming these risks and assigning an owner to each one. Confirm with counsel or your professional body how record retention and client consent apply in your jurisdiction before the agent touches live files.

How Be Human's managed approach handles this

Be Human begins with one role, such as the intake and reminder agent for a defined set of engagements. The first engagement sets the memory the agent may use, such as the firm's checklists and calendar, the tools it may call, the approvals staff must give, and one success criterion the firm can observe. Managed operations then cover monitoring of the communication log, recurring evaluations of draft accuracy, permission review when staff roles change, and controlled updates to checklists. Filing and judgment work is not part of that role, and the boundaries stay in writing so the firm can check them.

Choosing the first task

Pick one recurring request or one deadline your team already tracks by hand. Write down who approves it, which records the agent may read, and what the agent may never do. That short note is the start of a role that a reviewer can check. If you want a specific answer for your firm, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent prepare tax returns for a small CPA firm? Not as an autonomous step. The agent can gather documents, organize the file, and draft a summary for a preparer, but preparation, review, and any signature remain with licensed staff. Treat any output that could enter a return as a draft that requires professional review.

Is it safe to let an agent send deadline reminders to clients? It can be, provided every batch passes through an approval step and each reminder names only items that are actually missing. Start with one reminder type, review the drafts over several cycles, and only then consider narrowing the approval. Keep approval mandatory for anything that mentions a change to a deadline.

What should a firm document before setting up an agent? Write down the request types in scope, the staff member who approves each one, the sources the agent may read, and the tasks it must never perform. A short document like this makes the first engagement easier to scope and gives reviewers a standard to check against later.

How do we know the agent is doing its job? Define one observable criterion before setup, such as whether each reminder matches the open items in the file at send time, and check a sample on a regular schedule. Log corrections and use them to update checklists and instructions. If the agent cannot meet the criterion, narrow its role rather than adding permissions.

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