Skip to content
be-human.ai
← Back to the journal

AI by industry

Superintelligence Prep for Small CPA Firms: Document Requests

A small CPA firm can give an AI agent governed memory for document requests and filing status, while licensed accountants keep judgment and sign-off.

Be Human typographic cover with the title: Superintelligence Prep for Small CPA Firms: Document Requests
Conceptual illustration created with AI · Be Human

An AI agent for a small CPA firm can keep a governed record of document requests, client replies and filing status, so the team stops rebuilding the same chase from email threads and portal logs. The agent collects, sorts, reminds and drafts. Licensed accountants keep every judgment about what a document means, which tax position applies, and what is signed, filed or sent.

Key takeaways

  • Document requests and filing status are a strong first memory for a CPA firm because the work is repetitive, traceable and easy to check against a standard checklist.
  • The agent may track, classify, remind and draft. It should not decide taxability, positions, adjustments, extensions or penalty exposure without a licensed accountant.
  • Every stored fact needs a source, a capture time and a named owner, so the team can tell a confirmed item from an assumption.
  • Read access to client files and write access to the tracker should be separate permissions, with no path from the agent to filing or transmission.
  • A first engagement works best with one role, one checklist and one observable criterion, such as the number of open requests without a named owner, reviewed every week.

What a document request tracker needs to remember

Small firms usually run on a mix of email, portal uploads, spreadsheets and phone notes. A request for a brokerage statement, a prior-year return or a payroll summary may be sent three times, answered in part and then forgotten. The useful memory is not the whole client file. It is a compact ledger that records which item was requested, under which engagement, from which person, through which channel, what has arrived, what is still missing, and which accountant owns the next step.

Firms that already maintain a standard request list have an easier start, because the agent can work from a list the team already trusts. The approach described in AI for CPA firms: document requests and filing checklists follows this logic. Document-heavy operations in other sectors face the same pattern of partial replies and scattered status, as in AI agents for import and export companies.

Where the agent's memory stops and professional judgment begins

The boundary matters more than the software. An agent can record that a file arrived and that it resembles the item requested. It cannot decide that the file is complete, relevant or sufficient for the return, because that is a professional conclusion. Keep these decisions with licensed accountants:

  • Whether a document is complete, legible and sufficient for the position the firm intends to take
  • How income, deductions, credits, basis and dependents are characterized
  • Whether to request an extension, file late or accept a penalty or interest exposure
  • Client advice, scope changes, conflict checks and engagement terms
  • Anything signed, attested, filed or transmitted to a tax authority or third party

The same separation appears in evidence-heavy practices. In AI for property tax appeal firms, the useful work is organizing assessments, evidence and deadlines, while the argument itself stays with the professional. A CPA firm should draw the line in the same way, and write it down before the agent touches a live client.

A step-by-step plan for the first engagement

A first engagement should be narrow enough to review line by line. The sequence below keeps the scope small and the approvals visible.

  • Step 1: Export the firm's standard request list and mark which items the accountants treat as required for each engagement type.
  • Step 2: Choose one role, such as intake coordinator for individual returns, and write down what it may read, write and send.
  • Step 3: Define the memory fields: client, engagement, item, requested date, channel, received status, owner, and the reviewer who confirmed completeness.
  • Step 4: Set approval gates. Any client message that mentions tax treatment, any change to a deadline, and anything leaving the firm goes to a named accountant first.
  • Step 5: Run the role in shadow mode for a few weeks. The agent drafts reminders and status updates, and staff compare them with what was actually sent and why.
  • Step 6: Review errors weekly, update the checklist and the permission scope, and widen the role only after the errors stop repeating.

An illustrative scenario: four missing items before a deadline

Illustrative scenario, not a client case: a small firm prepares personal returns for several households. One household has a missing mortgage interest statement, a second brokerage form, a childcare receipt and an unsigned engagement letter. The agent reads the request ledger and the portal upload log, marks the first two items as outstanding, and notices that the childcare receipt has a reply that raises a dependent question. It does not interpret that question. It routes the thread to the assigned preparer with the reply attached, and drafts a reminder that lists only the three remaining documents and the signature. The preparer decides whether the receipt supports the claim, and an accountant approves the wording before anything is sent. At the end of the day, the tracker shows each open item with an owner and a last-touched time, which is the observable result the firm can check.

Permissions, data handling and what the agent must never touch

Treat the agent as a software identity with its own permissions, not as an extension of a staff login. Grant only the scopes the role needs, and review them on a schedule. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) describes the general principle: each action should be checked against what that identity is allowed to do, not against what the system can technically reach.

Client documents also carry a prompt injection risk. A scanned letter or an email can contain text written to change an agent's behavior. Treat document contents and client messages as data to be summarized, never as instructions to follow. The [OWASP prompt injection prevention cheat sheet](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) lists the controls that matter here, including separating instructions from untrusted content and limiting what a model can trigger.

In practice, a sound boundary for this role looks like this:

  • Read access to the request ledger and the portal index, not the full document store by default
  • Write access limited to the tracker and to draft messages placed in a review queue
  • No permission to change filing status in tax software, to submit anything to an authority, or to message a client without approval
  • Retention and deletion rules set by the firm and aligned with the personal data obligations that apply to it

How Be Human's managed approach handles this

Be Human designs the first engagement around one role, the memory and tools that role needs, the approval boundaries and an observable success criterion. The memory stays under the firm's control, with sources recorded for each fact. After launch, Be Human can run monitoring, recurring evaluations on past request threads, permission reviews, integration maintenance, incident response and controlled changes to the checklist or the prompts. The licensed accountants remain the people who decide what the documents mean and what leaves the firm.

How to tell whether the agent is doing its job

Every action the agent takes should leave a trace: what it read, what it wrote, which rule it followed and who approved the output. Standard observability concepts, such as logs, traces and metrics tied to a single workflow, are a good model for this, as described in the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/). Keep the measures simple and operational. Count open requests without a named owner. Count reminders that accountants edit before sending, since frequent edits point to weak drafting. Count misclassified items found during the weekly review. Run a fixed set of past request threads through the agent after each change, so a new prompt or checklist does not quietly break a working step. If your firm wants to find the first task worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.

Frequently asked questions

Can an AI agent prepare or file tax returns for a small CPA firm? The agent can organize requests, track received items, draft reminders and prepare summaries for a preparer. Preparing the return, deciding positions and filing or transmitting anything remain professional and approval-based tasks for licensed accountants, so the agent should have no direct filing permission.

Who owns the client documents the agent reads? The firm owns its client files and sets the rules for access, retention and deletion. The agent should work only inside the scopes the firm approves, and those scopes should be reviewed on a schedule and recorded, so the firm can show who had access to what and when.

How does the agent know that a document is the right one? It can compare a file's name, metadata and the reply text against the requested item, and flag a likely match. It does not decide that the document is complete or sufficient. A named accountant confirms completeness, and that confirmation is stored in the memory as the source of the status.

Where should a firm start? Start with one narrow role, such as intake coordination for one engagement type, using the firm's existing checklist. Run it in shadow mode, review errors every week, and widen the role only after the same errors stop recurring. The first useful task is usually the one with the clearest checklist and the most repetitive follow-up.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

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

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →