AI for Tax Preparation Offices: Chasing Documents Without Touching Figures
A practical guide to an AI agent that chases missing tax documents during peak weeks, keeps return figures out of its reach and builds supervised memory.

An AI agent for a tax preparation office can chase missing client documents during peak weeks by sending approved requests, tracking what arrives, and escalating gaps to a named preparer. It works from a checklist for each client and writes only to the document tracker and approved message templates. Done this way, the agent removes hours of follow-up from the busiest weeks without changing a single figure on a return.
Key takeaways
- Document chasing is the most repetitive load in filing season, and it is the safest place to start with an agent.
- The agent owns requests, reminders and status. Income, deductions, credits, classifications and filing decisions stay with licensed preparers.
- Every request needs a named document type, a client, a due point and an owner who can close it.
- Permissions should be scoped to the tracker and approved messages, with read-only access wherever return software is involved.
- Each request, reply and escalation becomes a supervised record that a later agent can rely on.
Why document chasing consumes the busiest weeks
Preparers rarely lose time on the tax work itself during peak weeks. They lose it on the missing piece: a brokerage form not yet issued, a childcare receipt sitting in a spouse's email, a prior-year letter nobody requested from the client. Each gap needs a message, a reminder, a note of what came back and a check that the item is really complete. Multiply that across a full client list and the follow-up becomes a second job that competes with the first.
Most offices track this work in a mix of spreadsheets, inboxes and memory. Items fall through because the calendar is full, not because anyone is careless. An agent is useful here because it is consistent. It sends the same clear request on the same schedule and records the outcome every time, which is the part people find hardest to sustain in February.
Where the agent's authority should stop
The boundary between what the agent may do and what it must never do needs to be written down before the first message goes out. The agent may:
- Read the document checklist for each client and the current status of every item.
- Send approved request templates in the client's language and preferred channel.
- Record receipt notes, such as a scan that is unreadable and needs to be requested again.
- Escalate to the named preparer when a document conflicts with the prior year or appears incomplete.
The agent must not enter, edit or infer any figure on a return. It must not decide how a document is treated for tax purposes, advise on eligibility, send messages outside approved templates or promise a filing date. If the client asks whether a deduction applies, the agent hands the question to a preparer and records that it did so.
This is least privilege applied to a real workflow. The OWASP authorization guidance treats access control as something the system enforces, not something a well-behaved assistant is trusted to respect. For a document agent, that means the credentials cannot write to return fields at all, so the boundary does not depend on the model's judgment.
A step-by-step plan for the first engagement
A first engagement should be narrow enough to finish and measure. A workable sequence looks like this:
- Step one: list the document types that generated the most follow-up last season, taken from the office's own tracker rather than from memory.
- Step two: write one approved template per document type, covering the client's language, the deadline wording and the exact list of what to send.
- Step three: define escalation rules, including which mismatches go to which preparer and how quickly they should be reviewed.
- Step four: connect the agent to the tracker and the messaging channel with scoped credentials, and log every action it takes.
- Step five: run a supervised period in which a staff member approves each outgoing message before it is sent.
- Step six: set a success criterion that can be observed, such as the share of open requests closed before the internal review date, measured each week.
- Step seven: review the logs after the peak weeks and decide, item type by item type, what the agent may send without approval.
Illustrative scenario: a missing account statement two weeks before review
Illustrative scenario, not a client case: a household file lists a savings account in the prior year, but this season's documents include no statement for it. The agent sends the approved request in the client's preferred language, logs the message and schedules a reminder. The client replies that the account was closed during the year. The agent does not decide what that closure means for the return. It records the reply, flags the mismatch with the prior-year file to the preparer and marks the item as needing a human answer. The preparer decides how the closure affects the return. The log shows who was asked, when, and what came back, so the preparer does not have to reconstruct the conversation.
How supervised memory prepares the office for later agents
Each request, reply, escalation and approval is a record. If those records carry clear fields, including client, document type, date requested, date received, channel, approver and outcome, they become the memory a future agent can work from. Retrieval from approved records is more reliable than asking a model to recall what happened last year. The NIST AI Risk Management Framework emphasizes documenting the context, risks and governance of an AI system across its lifecycle, and an agent's activity log is a practical place to begin that work.
Keep the memory bounded. Anthropic's guidance on context engineering describes keeping an agent's working context focused on what the current task needs. A document agent needs the checklist, the template and the open items, not the full history of every client's tax life. Store document status and communication history as memory, and keep tax computations out of it unless a human reviewer owns that process. An office that separates these early will find the later layers easier to govern.
For a parallel in another service business, the same discipline of tracking deadlines and evidence appears in AI for Property Tax Appeal Firms, where assessments, evidence and deadlines are the working memory.
How Be Human runs this as a managed service
Be Human designs the first engagement around one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. In managed operation, Be Human can monitor the agent's activity, run recurring evaluations of message quality and escalation accuracy, maintain the tracker and messaging integrations, review permissions before each season and respond to incidents such as a message sent to the wrong client. Changes to templates and boundaries are made in controlled steps, with the office approving each one. The service is delivered remotely in English and Portuguese, and the office keeps decisions about returns with its own preparers.
Be Human's approach to follow-up work is similar in other sectors where records must be complete before a human can act, as described in Superintelligence Prep for Medical Billing Firms: Denials and Follow-Up.
Choosing the first task for your office
The strongest first task usually combines high repetition, low judgment and a clear owner. Document follow-up in a tax office often fits that profile, but only if the office can name who closes each document type and where the audit trail will live. Start with the document types that cause the most back-and-forth, and leave anything that touches a figure for a later decision, if it ever comes. If you want to find the first task worth handing to an agent in your office, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent prepare tax returns for the office? In the model described here, it does not. The agent handles document requests, status tracking and escalation, while figures, tax treatment and filing decisions remain with licensed preparers. Local professional rules should be confirmed with the office's own advisers before any scope is set.
What data does the agent need to access? It needs only the document tracker fields and the approved messaging channels, using scoped credentials that cannot write to return software. Handling of client personal data should follow the privacy law that applies to the office and to its clients, and that review belongs in the first engagement, not after it.
How is the first engagement scoped? It is scoped to one role, a short list of document types, a set of approved templates and one observable success criterion. The narrower the first scope, the easier it is to review the logs, adjust the boundaries and decide what the agent can do next without approval.
What happens when a client replies in the wrong language or sends an unreadable scan? The agent uses templates in each language the office serves and records a receipt note when a scan cannot be read, then requests a clearer copy. Anything ambiguous, such as a reply that changes the meaning of a prior-year item, is escalated to a named preparer rather than resolved by the agent.
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.


