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

AI by industry

Superintelligent Agents for Tax Offices: Document Intake in Season

Agents can run seasonal document intake, chase missing forms and keep client status current. Filing positions and sign-off stay with the enrolled preparer.

Be Human typographic cover with the title: Superintelligent Agents for Tax Offices: Document Intake in Season
Conceptual illustration created with AI · Be Human

Superintelligent agents for tax offices are AI agents that collect client documents, track what is still missing, send approved reminders and keep a current status record for every file during the busy season. They do not decide how income, deductions or credits are treated, and they do not sign or file anything. The enrolled preparer keeps every filing decision, while the agent makes sure the preparer receives complete, clearly labeled files on time.

Key takeaways

  • The agent owns the document trail and the status record. The preparer owns every position taken on a return.
  • Keep a checklist memory for each client that shows what was received, what is pending and what was said, with dates.
  • Write approval boundaries before launch: which messages send automatically, which need staff review and which never send.
  • Judge the first season by complete files reaching preparers and fewer repeated chase cycles, not by message volume.
  • Treat privacy, access scope and client-uploaded text as design inputs from day one, not as cleanup after launch.

Why season intake breaks down in a tax office

In most tax offices the intake problem is not a lack of effort. Documents arrive by email, through upload links, as phone photos, by mail and sometimes as a forwarded message from a spouse or bookkeeper. Each arrival has to be matched to a client, a tax year and a checklist item, and that matching is usually done by the same people who prepare the returns. When the season peaks, the matching slips first, and a file can look complete in one inbox while a key statement sits unread in another.

The failure is usually about state rather than any single document. Nobody can say quickly which files in progress are waiting on a prior-year letter, or which client already answered a question two weeks ago. Document trails in other regulated work follow the same logic. In import and export document status, a missing certificate blocks a shipment until someone records it and moves it forward. Tax intake has the same shape, with higher stakes for each missing page.

What an intake agent should own

A useful first role is narrower than it sounds. The agent works from the firm's engagement checklists and file records, and its responsibilities are concrete:

  • Receive documents from approved channels and attach each one to the correct client and tax year.
  • Check every upload against the checklist for that client type, and mark items as received, unreadable or unrelated.
  • Send scheduled reminders for pending items, and record every message sent and every reply.
  • Maintain one status record per file that a preparer can read in under a minute.
  • Route any question the record cannot answer to a named staff member.

The agent does not read a return as a whole, does not infer what a document means for the tax position, and does not decide whether a missing item matters. It flags and files, and people judge.

A step-by-step plan for the first season

Start with one group of clients whose documents are most repeatable, usually returning clients who send the same set each year. Avoid unusual engagements for the first run, because they test judgment rather than process. A workable sequence looks like this:

  • Step 1: Choose one client group and one document set for the pilot.
  • Step 2: Write the checklist with the preparer, marking each item as required, conditional on the client's situation or optional.
  • Step 3: List the channels the agent may use, such as the firm's secure upload link and an approved sending address, and list the channels it may never use.
  • Step 4: Set the chase schedule, the maximum number of reminders and the point at which staff take over the conversation.
  • Step 5: Define the status fields: received, pending, questioned, ready for review and released to preparer.
  • Step 6: Run a shadow period in which staff compare the agent's status records with their own tracker each day.
  • Step 7: Review every miss weekly, tighten the checklist wording or reminder text, and only then expand to the next client group.

How to chase missing forms without damaging client trust

Chasing is where clients notice the agent most, so the messages should be short, specific and plain. Each reminder names the exact item, explains in one sentence why it is needed and points to the secure way to send it. Messages never speculate about penalties, never suggest what a client should report and never comment on a client's tax position. Reminders stop at the limit set in the plan, and any reply that reads like a question about tax treatment goes to a person.

Illustrative scenario: A client uploads a year-end bank statement for one account, but the prior-year file listed a second account that has no statement yet. The agent marks the second account as pending, sends the approved reminder asking whether that account was closed or still open, and tags the file for the preparer. If the client replies that the account was closed during the year, the agent does not work out what the closure means for the return. It attaches the reply to the file and the preparer decides what to do with it.

What the client status memory stores, and what it must not infer

The memory is what makes an agent useful across weeks rather than in a single exchange. For each file it should hold the checklist state, the dates documents were received, the history of reminders and replies, open questions with the staff member who owns each one, the preferred contact channel and the consents on record. It should answer questions such as what is still missing for a given client from that record, not from a general guess about what clients usually send.

Limits matter as much as storage. The memory should not hold conclusions the preparer has not confirmed, such as a guessed employment type or an assumed residence. Anything inferred should be labeled as unconfirmed, and sensitive identifiers should be kept only where the workflow needs them. Retention should follow the firm's own obligations. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for naming these risks and documenting how they are handled before launch.

Which filing decisions stay with an enrolled preparer

Be specific about the line, because the line is what protects the firm and its clients. The following decisions stay with the enrolled preparer, or with the credentialed professional your jurisdiction requires:

  • Whether a document supports a deduction, credit, exclusion or a particular income treatment.
  • Which form, schedule or election applies to the client.
  • How to respond to a notice from a tax authority, and whether to request more time.
  • Whether to amend, disclose or correct anything from an earlier year.
  • Signing, e-file authorization and release of any return to the client.
  • Engagement scope, fee discussions and any client request that changes the work.

The agent can still help around these decisions. It can log that a notice arrived and attach it to the file, and it can prepare a neutral summary of what the file contains, clearly marked as a summary and not a recommendation. It should never draft a position for a client to accept. A simple test helps: if the answer changes what the client owes or claims, a credentialed person decides it.

How a managed setup holds the boundary through a season

Be Human designs and operates this kind of agent as a managed service. The first engagement defines one intake role, the checklist memory it needs, the tools and channels it may use, the approval boundaries described above and an observable success criterion, such as how many files reach a preparer complete by an agreed checkpoint. Operations then cover monitoring, recurring evaluations of the agent's status records against staff trackers, permission reviews whenever staff or channels change, integration maintenance and incident response if a message ever reaches the wrong client. Improvements are made one change at a time, so each change can be traced to a stated reason.

Access control needs the same discipline. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) points toward least privilege, so the agent should see only the files of clients it is working on, use separate permissions for reading documents and sending messages, and write every action to an audit trail. Client-uploaded text also needs care. A scanned letter can contain instructions aimed at an AI system, so the agent should treat document content as data to file, never as commands to follow. The [OWASP Top 10 for LLM Applications](https://genai.owasp.org/llm-top-10/) lists prompt injection and excessive agency among the risks to plan for. If you want to find the first intake task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Can a superintelligent agent prepare or file tax returns? No. The agent collects, sorts, chases and records documents, while preparation, tax positions and filing stay with the enrolled preparer. The agent has no authority to sign or submit anything on the firm's behalf.

Which client messages can be automated? Routine reminders for items on the approved checklist can be automated after a shadow period, using wording the preparer has approved. Anything that asks about tax treatment, disputes a document or mentions a notice goes to a named staff member.

Where should client documents live? Documents should stay in the firm's existing secure storage, with the agent's access limited to the files it is working on. Retention and privacy duties depend on your jurisdiction, so confirm them with your own advisers before the agent reads any client file.

How do we know the agent is working well during the season? Compare its status records against staff trackers on a regular schedule, review every miss, and log each incident with its cause and the fix. Those reviews show whether the checklist, the reminders and the boundaries still match how the office works.

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 →