Superintelligence for Accounting Firms: Tax Season Document Chasing
A practical plan for accounting firms to automate tax season document chasing, while partners keep sign-off on every filing and every client message.

Superintelligence for accounting firms, in practical terms, means AI agents that take over the repetitive chase for client documents during tax season: requesting missing statements, tracking what has arrived, and flagging gaps to the staff who own each return. Partners keep sign-off on every filing, every change to a figure, and every commitment made to a client, while the agent works inside permissions that the firm sets and reviews.
Key takeaways
- Document chasing is a workflow problem with a clear owner, which makes it a sound first task for an agent.
- The agent should request, track, remind and match files. It should never decide what a document means or approve a return.
- Write the approval boundaries before the agent sends its first message, not after the first mistake.
- Keep one status record per client that both staff and the agent read from, so open items never live only in an inbox.
- Judge the work by whether open requests reach a preparer with complete information, and whether clients receive only the reminders they actually need.
Why document chasing consumes the season
During tax season the bottleneck is rarely the tax work itself. It is the wait for a bank statement, a corrected brokerage form, a signed engagement letter, or a property record the client was sure was in a folder somewhere. Staff spend hours writing the same request in slightly different words, opening attachments to learn what was actually sent, and rebuilding a list of what is still missing from memory or from a spreadsheet nobody fully trusts.
The cost shows up later as delay. Returns stall at intake, reviewers receive incomplete files, and partners finish the season with open questions about who promised what and when. Firms that run other document-heavy work already know that the missing item is usually the one that stops the file. The same pattern appears in property tax appeal firms, where evidence and deadlines have to be tracked together or the appeal weakens.
Decide what the agent may touch
Before naming any tool, sort the work into three groups. The sorting is the most important design decision in the project, and it should be written down and signed off by a partner.
- The agent may act alone: send scheduled reminders from approved templates, log received files against each client checklist, rename and classify uploads, and answer "what do you still need from me" with the checklist text.
- The agent may draft, and staff must approve: wording for unusual requests, any message that mentions a deadline, penalty or extension, and replies to questions about a figure from a prior year.
- Only a person may act: interpret a document, change a number in a working paper, decide whether an item qualifies for a treatment, approve or transmit a filing, waive a missing item, or promise a date to the client.
This split keeps the agent useful in the volume work while leaving judgment where accountability already sits.
A step-by-step plan for the first tax season
Start with the checklists, because the agent can only chase what the firm has defined. Then build the workflow around them.
- Define a fixed document checklist for each client type, such as individuals, sole traders, small partnerships and payroll-only clients.
- Create one client record for each engagement with the owner, the expected documents, the status of each item, and the date each request was sent.
- Connect the agent to the mailbox or portal clients already use. Give it read access to uploads and write access only to the request log and the approved templates.
- Write the reminder sequence and the escalation rule. After the agreed number of unanswered reminders, the file goes to the named staff member with the open items listed.
- Run the agent in draft mode for the first week, with a staff member reviewing every outgoing message before it leaves.
- Review the log at a fixed weekly meeting: which requests stalled, which templates confused clients, and which files arrived in the wrong format.
- Only after the draft review is clean, let routine reminders send automatically, and keep escalations under human control.
The same discipline of one shared status record is what makes the work traceable. The logic is close to the document flow in import and export firms, where the status of each file matters more than any single message.
What a chase looks like in practice
Illustrative scenario, not a real client account: a household has uploaded two of five expected documents. The agent sees the missing brokerage statement on the checklist and sends the approved reminder, which names the exact statement needed and the portal location where it can be uploaded. The client replies with a scanned form, but the year printed on it does not match the return being prepared.
The agent does not treat the form as the correct document. It logs the file as received with the year unclear, tells the client that a staff member will confirm it, and places the file in the preparer's queue with the question written out in one sentence. The preparer checks the form, decides whether it is usable, and updates the record. The agent handled the chase and the paperwork trail, and the judgment stayed with the person who owns the return.
Where partners keep sign-off
The agent can ask for a document. It cannot decide what the document means.
Sign-off works best as a gate rather than a courtesy. Every return has a named approver, and the agent's outputs are inputs to that review: a complete or incomplete status, a list of open questions, and any draft message. When a figure changes because a document arrived, the change should appear as an explicit difference that the preparer accepts, not as a silent update to a working paper.
Client commitments need a rule of their own. A promised filing date, a note about an extension, or any statement that touches a penalty comes from staff. If a client disputes what the agent said, the agent stops, records the exchange, and routes the conversation to a person. Partners who review the log weekly will see these routing decisions and can tighten the rules where clients push.
Access, privacy and the failure modes to plan for
Tax files contain identifiers, income details and family information, so the agent should receive the least access it needs. The OWASP authorization guidance is a useful reference here: enforce access checks on the server side, scope each tool to the client files it actually needs, and do not rely on instructions in a prompt to keep one client's data away from another. Where personal data is handled this way, privacy rules such as Brazil's LGPD apply, so the firm should keep a plain record of what the agent processes and why.
Client documents are also untrusted input. A scanned letter or an email attachment can contain text written to steer an AI system, for example instructions to change a request or to reveal another client's details. The OWASP prompt injection prevention cheat sheet explains why the agent must treat document contents as data to be read, never as instructions to be followed. The NIST AI Risk Management Framework offers a structure for identifying these risks, measuring them, and assigning an owner before the first season begins. Common failure modes to plan for include a file matched to the wrong client, a reminder sent with the wrong template, a duplicate reminder after a reply, and a document read as the wrong year. Each one should have a written rule and a person who receives the alert.
Running it after the first season
Be Human designs the first engagement around one useful role, in this case the document-chasing agent. The design defines the checklist memory the agent needs, the tools it may call, the approval boundaries described above, and one observable success measure, such as the share of open requests that reach a preparer with complete information. Managed operations then cover monitoring of the request log, recurring evaluations of message quality against the approved templates, maintenance when a client portal or mailbox integration changes, periodic permission reviews, and incident response when something goes wrong. Template and rule changes are scheduled between filing periods rather than during the busiest weeks, so the firm is never learning a new process in the middle of a deadline. 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
Does the agent need access to completed tax returns? Usually it does not. For most of the chase it needs the client checklist, the upload log, and the approved templates, and it should not open finished returns at all. Limit access to the files that each task requires, and review that scope before each season.
Who is responsible if the agent sends a wrong reminder? The firm remains responsible, which is why every outgoing template is approved by a named person and the first week runs in draft mode. Each message is logged, so the error can be traced, corrected with the client, and the rule that caused it can be changed.
Can clients reply in Portuguese or English? Yes, provided the templates and escalation wording have been approved in each language. The agent should use the language the client has used before, and any reply that is ambiguous in either language goes to staff rather than being guessed.
How is this different from a shared inbox with a reminder tool? A reminder tool sends messages on a schedule but does not know what has arrived, what is still missing, or who owns the next step. The agent reads and writes one shared status record, classifies what comes back, and escalates with the open questions written out, while the approval rules keep judgment with the partners.
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.


