Superintelligence Prep for CPA Firms: Client Documents and Deadlines

An AI agent for an accounting firm works best when it takes on two narrow, checkable jobs: collecting client documents against a fixed checklist, and tracking deadlines for each client entity. Both jobs produce a running record of what each client sent, when it arrived, and what is still missing, and that record becomes the firm's most useful asset as models change. Be Human designs these agents so that staff approve anything that reaches a client or changes the status of a file.

Key takeaways

• Start with one engagement type and one document checklist, not the whole practice.
• Keep client memory in a structured index you control: documents, dates, open requests and owners.
• Give each agent read access first, and require staff approval for every outbound message and status change.
• Treat text inside client files as untrusted input, because a scanned letter or an email can contain instructions.
• Measure the agent against past engagements with known gaps, and repeat that check whenever the model or workflow changes.

Why document collection is the right first agent

Most of the calendar pressure in a small CPA firm comes from the same chase repeated across many clients. A request goes out, a reminder follows, a partial set arrives, someone sorts it, and the deadline gets closer while one statement is still missing. The work is repetitive, the rules are already written in engagement letters and checklists, and a missed item is costly enough that people want a clear record. Those properties make it a good candidate for an agent, because the agent can compare what has arrived against a known list without exercising professional judgment about the tax position itself.

The deadline side has the same shape. Each client entity has its own filing dates, extension rules and responsible person. A spreadsheet works until a partner changes a date, a client adds an entity, or a staff member goes on leave. An agent that reads the same calendar every morning, lists what falls due in the next stretch, and names the person who owns each item keeps dates from living only in someone's memory.

Client memory that outlasts any single model

Preparing for more capable models is less glamorous than it sounds. It means arranging what the firm knows so that a better model can use it on the first day. General capability changes over time. What the firm knows about each client does not: which documents were received, which were promised, which questions were answered and by whom. If that knowledge lives in structured records and files the firm controls, the agent retrieves what it needs at the moment of a task. If it lives only in a chat history or in a tool's hidden settings, switching to another model means rebuilding it.

Retrieval is the practical pattern here. When the agent drafts a request, it pulls the relevant checklist, the client's document index and the last few entries in the file, instead of relying on facts the model is assumed to remember. The AWS guidance on comparing retrieval with fine-tuning explains why this approach is usually easier to update when facts change. For a firm, a new document rule becomes an edit to a record rather than a retraining project.

Illustrative scenario: a firm with one engagement type keeps a one-page checklist for each client, covering bank statements, payroll summaries and prior-year returns. When a client's bank changes its statement format, staff update the checklist once. The next request the agent drafts reflects the new format, while the index still shows which months were received under the old one.

Firms that already run intake for legal or mortgage work face the same pattern. AI Agents for Law Firms: Setup Guide for Intake and Deadlines and AI for Mortgage Brokers: Documents, Updates and Reminders show how similar document indexes and reminder rules are set up in those settings.

How to design two small agents with clear limits

Keep each agent small enough that a reviewer can describe its job in one sentence. The document-collection agent reads the engagement checklist, compares it with the files received, drafts a request for anything missing, and logs each receipt with its date and source. The deadline agent reads the filing calendar, lists items due within a set window, and drafts a reminder for the named owner. Neither agent decides whether a return is complete, whether a deduction is allowed, or what a client should be told about their tax position. Those remain professional decisions.

Use this guide to sort each action the agent could take:

• If the action only reads records or compares lists, the agent may run it without review.
• If the action drafts a message to a client, the agent may write the draft, but a staff member approves it before it is sent.
• If the action changes a status, such as marking a document received or a filing ready, the agent may propose the change, but a named person confirms it.
• If the action involves payment, signatures or submission to an authority, the agent does not perform it.

Write these rules down before the first run. A limit that exists only in someone's head tends to disappear when the agent produces a helpful-looking shortcut.

Keeping staff in the approval path

Human review only works when it is specific. A reviewer shown a vague summary will approve whatever looks reasonable. Each approval item should show the full draft, the checklist line it relates to, the record that triggered it, and what will change if it is approved. Schedule reviews at fixed times, such as a short daily pass over requests and a weekly pass over the deadline list, so approvals do not pile up or get skipped.

Access should match the job. Give each agent its own identity, with permission to read the document index and the calendar, and to write only to drafts and logs. OAuth is the common standard for granting a tool limited access to an account without sharing passwords. The OWASP authorization guidance describes checking permissions on every action rather than once at setup. Remove access when an engagement closes or a staff member changes role.

Checklist for the first deployment:

• Each agent has a named person inside the firm who reviews its output.
• Every outbound message waits in a draft queue until it is approved.
• Every change to a document status is logged with the agent, the source record and the approver.
• Access is limited to the specific folders and calendars the job needs.
• A person can pause either agent in one step.

Treating client files as untrusted input

Client documents come from outside the firm, which makes them a risk surface. A forwarded email, a scanned letter or a spreadsheet can contain text written to steer an automated reader, whether by accident or on purpose. The OWASP guidance on prompt injection prevention describes this class of problem and the controls that reduce it. For a document agent, the practical rules are simple. Extract specific fields, such as statement period and account type, into the index, rather than letting free text drive actions. Never let text inside a file change what the agent is allowed to do. Show the reviewer the page or line that supports each entry, so a mistake is visible before it is acted on.

Running it as a managed service and measuring the result

A first deployment usually fails quietly. Requests go out late, a checklist drifts from how the firm actually works, or an agent keeps proposing a status that reviewers always reject. Be Human runs these agents as a managed service. The first engagement defines the one role, the checklists and calendar it reads, the approval boundaries and an observable success criterion, such as the number of open items older than an agreed age. After launch, monitoring, recurring evaluations, permission reviews and controlled changes to prompts or checklists keep the agent aligned with the firm's process. Incidents, such as a request sent with the wrong client name, follow a written response path, and each change is reviewed before it goes live.

Evaluation is the part most firms skip. Choose a set of past engagements where you already know which documents were missing and which deadlines were close. Run the agent against those records and compare its output with what staff recorded at the time. Repeat the same test whenever the model, the checklist or the calendar source changes, because a result that held before an update may not hold after it. The NIST AI Risk Management Framework is a useful reference for assigning an owner, a control and a review step to each of these risks.

Success should be concrete. For the document agent, a useful criterion is that every drafted request maps to a checklist line, and that a reviewer can trace each receipt to a file. For the deadline agent, it is that no item in the coming window lacks a named owner. Both can be checked without judging the quality of the tax work itself. If you want to find the first task in your firm worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can an agent send document requests to clients on its own? No. In the design described here, the agent drafts each request and places it in a queue, and a named staff member approves it before anything reaches the client. This keeps the firm's voice and the engagement terms under human control, and it gives the reviewer a chance to catch a wrong name or an outdated checklist line. Automatic sending can be reconsidered later, but only after the firm has a documented review record.

Do we have to retrain anything when a better model arrives? Usually not. The checklists, document index and calendar sit outside the model and are retrieved at the moment of each task, so a newer model can use the same records. You should still rerun the evaluation set, because a more capable model may behave differently on the same prompts. Plan that test before you switch.

Where does client data live, and who can see it? The design keeps client records in storage the firm controls, and each agent receives scoped access to specific folders and calendars. Duties on confidentiality and data protection depend on your jurisdiction and your client agreements, so review them with your advisers before connecting any system. Permission reviews are part of the managed service so that access does not widen quietly over time.

How do we know the agent is reliable enough to trust? Do not decide by impression. Run the agent against past engagements, compare its output with what staff recorded, and agree on the threshold before launch. Keep the threshold narrow, such as every drafted request mapping to a checklist line, and review each failure with the team. Trust grows when the record of approvals and corrections shows the same pattern over time.

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