AI Agents for Tax Preparation Offices: Document Chasing and Extensions
How tax preparation offices can use AI agents for document requests and extension reminders while preparers keep sign-off on every return. Scope, approvals and checks.

An AI agent for a tax preparation office can chase missing client documents, send scheduled requests and reminders, and tell the team when an extension or filing window is closing, all within rules the office sets. Preparers remain responsible for every return: the agent gathers, tracks and drafts, while a credentialed preparer reviews and approves each file before anything is signed or submitted. The practical starting point is one narrow workflow, the document request cycle, with clear approval boundaries and a success check the office reviews every month.
Key takeaways
- Document chasing is a strong first job because it is repetitive, rule-based, and fails quietly when nobody follows up.
- Preparers keep sign-off on every return, every figure sent to a tax authority, and every filing. The agent holds no authority to submit anything.
- Extension reminders should come from a calendar the office owns and maintains, with each message logged so the team can show who was told what and when.
- Give the agent only the memory, file access, and tools its single role needs, with read access by default and any write action gated by a person.
- Judge the workflow on one observable criterion, such as how many open requests are resolved or escalated before the weekly review, and inspect failures every month.
Why document chasing is the right first job
Much of the pressure in a tax office comes from waiting. A preparer can be ready to work on a return and still be blocked because a client has not sent an income statement, the prior-year file, receipts for a rental property, or the identity documents a new client was asked for weeks earlier. Chasing those items is a long chain of emails, calls, and small follow-ups, and it often depends on one person remembering to check the folder again on Thursday.
That repetitiveness is what makes the task suitable for an agent. The rules are explicit: a checklist exists for each client type, the trigger is a missing item or a silent client, and the escalation path is known. The work involves little tax judgment, but it does involve personal financial data, so the first design decision is scope rather than model choice. The agent should see the request list and the document folders for the clients in the workflow, and nothing else.
Map the request cycle before you automate it
Before anyone configures an agent, write the cycle down as it works today. Most offices find that the process lives in a few people's habits rather than in any document. The map should answer these questions in order:
- List each document type by client category, such as employee, self-employed, business owner, and property owner.
- Define the trigger for each request: an engagement signed, an item missing at intake, or a client silent past an interval the preparer sets.
- Name the owner of each exception, whether the preparer, the front desk, or the client, so nothing falls into the gap between them.
- Decide which messages may go out from approved templates and which need a preparer to write or approve the wording first.
- Define the closing event: the document arrived, a person checked it, and the request is marked complete in the client file.
Once the map is written, the agent's job is to carry out the steps the office has already agreed on. If a step cannot be written down, it is not ready to automate, and it stays with a person for now. Offices with similar document-heavy tracking can compare their map with the pattern in AI Agents for Import Export Companies: Documents and Status.
What the agent may do, and what it may not
A plain decision guide, reviewed by the preparers themselves, is the most useful way to set boundaries. For document requests, a sensible scope looks like this:
- May send scheduled reminders from approved templates, with the client name and missing item taken from the request list.
- May log every reply, attach received files to the correct client folder, and mark items as received or unclear.
- May draft a short summary of what is still missing for each file, so the preparer starts the day with a list rather than a search.
- May route unclear replies, mismatched files, or any message that mentions a changed deadline to a named preparer.
- Must not interpret tax law, decide whether an expense is deductible, change any figure, sign anything, or submit anything to a tax authority.
- Must not send messages outside the approved templates without human review.
Incoming documents need particular care. A client may send a scanned file or an email containing text meant to change the agent's behavior, whether by accident or on purpose. The OWASP guidance on prompt injection treats external content as untrusted input. In practice, the agent should read documents as data to be filed and summarized, never as instructions, and its actions should stay within the same boundaries no matter what a file says.
Extension reminders need a calendar the office owns
Extension reminders are where a missed date becomes expensive. The rule that protects the office is simple: deadlines come from a calendar the office owns and checks, not from anything the agent remembers. The agent reads that calendar, sends reminders at intervals the preparers choose, and lists clients who have not responded before the next internal checkpoint. Deciding which deadlines apply to which client, and whether an extension is needed, remains a professional decision made by the preparer under the rules that govern the office. Offices that track assessment evidence and deadlines face a similar problem, described in AI for Property Tax Appeal Firms: Assessments, Evidence, Deadlines.
Every reminder and every reply should be recorded as an event with a timestamp, the template used, the client file it belongs to, and the person who was escalated to. This is the same discipline operations teams apply to software systems, and it matters here for a different reason: when a client says the notice never arrived, the office can show exactly what was sent.
Illustrative scenario: a small office with a few preparers runs the agent against its open-request list every weekday morning. The agent finds three self-employed clients who have not returned expense records, sends each the approved reminder, and routes a reply from a fourth client, which mentions a missing bank statement and a possible change of address, to the preparer who owns that file. The preparer reads the message, updates the file, and decides the next step personally. The weekly meeting starts with a list of exceptions instead of a search through inboxes.
Keep preparer sign-off on every return
Sign-off should be a state change in the workflow, not a habit. A return moves to review only when the named preparer marks it so, and it moves toward submission only through that preparer's own credentials in the tax software the office already uses. The agent should have no path to those states. Its permissions should be role-specific: read the request list and the folders for assigned clients, write to the communication log, and nothing else.
Access control deserves the same care as the workflow. The OWASP authorization guidance favors checking permissions on every access and denying by default, which translates here into per-client scoping and regular permission reviews. The NIST AI Risk Management Framework offers useful vocabulary for the governance side: who is accountable for the system, how it is tested, and what happens when it fails. A small office does not need a large program, but it does need a named owner, a review date, and a written way to pause the agent.
How Be Human runs this as a managed service
Be Human starts with one defined role, usually the document request cycle, and specifies the memory and tools that role needs, the approval boundaries, and a success criterion the office can observe. Managed operations then cover monitoring of the request queue, recurring evaluations of message accuracy and template use, integration maintenance when the email, calendar, or document portal changes, periodic permission reviews, incident response if a message reaches the wrong client, and controlled improvements to the workflow. The service is delivered remotely, in English and Portuguese.
How to measure whether it is working
Choose one criterion before launch and review it every month. Useful examples include the number of open requests resolved or escalated before the weekly meeting, the number of messages sent to the wrong client, and how quickly unclear replies reach a preparer. Build a small test set from real but anonymized replies, including messy ones, and check how the agent classifies each. Evaluation guidance for agents, such as Anthropic's notes on demystifying evals, makes the same point: test before changing, and keep the test set current. A monthly review of failures is more useful than a dashboard, because each failure points to a template, a rule, or a permission that needs adjusting. If you want to find the first task in your office worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent prepare or file a tax return? In the model described here, no. The agent collects documents, tracks open items, sends approved reminders, and organizes the file so a preparer can work faster. Preparation, review, sign-off, and submission stay with credentialed people using their own systems and credentials, and the agent has no permission to reach those steps.
Do we have to replace our tax software to start? Usually not. A first workflow works alongside the office's current email, calendar, and client folders, with access limited to what the request cycle needs. Integrations are part of managed operations, so a change to a portal or mailbox is handled as maintenance rather than as a surprise.
How do we protect client documents when an agent reads them? Scope access per client, deny by default, log every action, and set retention rules that match the office's obligations. Treat anything inside a received file as data rather than instructions, since documents can carry text that tries to change behavior. Review permissions on a fixed schedule and keep a written way to pause the agent.
What should a tax office measure in the first month? Track one criterion chosen before launch, such as how many open requests were resolved or escalated before the weekly review. Record messages sent to the wrong client as a separate incident, because they matter more than speed. Review failures with the preparers and adjust templates or rules first, then the workflow.
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.


