Superintelligence for Accounting Firms: What to Delegate and What to Keep

Superintelligence for Accounting Firms: What to Delegate and What to Keep

Which checks, document follow-ups and client messages an AI agent can handle in an accounting firm, and which need licensed sign-off.

•

Want to apply this in your company?

Book a call with Juan →

In the sense that matters to an accounting firm, superintelligence does not mean a system that thinks like the accountant and takes responsibility in their place. What an AI agent does well is run repetitive, verifiable tasks with written rules: check documents against a checklist, chase outstanding items on an approved schedule, and answer routine questions based on validated material. Signatures, guidance on tax treatment, and any opinion that binds the firm remain with the qualified professional.

Key takeaways

  • Delegate the mechanical checks: whether required fields are present, whether the period is correct, whether the parties are identified, and whether figures match closed rules.
  • Document follow-up works well as a routine: reminders, a record of each reply, and a handoff to the team when something falls outside the expected path.
  • Client communication needs a fixed scope: status updates, instructions for sending files, and questions the firm has already answered stay with the agent, while exceptions go to a person.
  • Anything that creates an obligation, an opinion, an effective ledger entry, or a signature requires review and technical responsibility from the accountant, without exception for convenience.
  • Start with one routine, an agreed success criterion, and decision records before expanding the scope. Firms that map the process first tend to make fewer mistakes when they scale up.

What separates a delegable task from one that needs technical responsibility

The useful question is not whether a task is important. It is three other questions: can the result be checked against a written rule? Is the error easy to reverse? Does the decision depend on judging a specific client's situation? If the first two answers are yes and the third is no, the task is a good candidate for delegation. If the decision creates an obligation toward the tax authority, the client, or third parties, the agent can prepare the material but cannot complete the action.

Anthropic's published guidance on effective agents distinguishes workflows, where steps are defined by code, from agents that decide their own next steps, and recommends starting with the simplest solution that solves the problem. For a firm, this turns into a practical rule: the more fixed the rule, the more freedom the agent can be given.

  • Can delegate: tasks with a written rule, checkable input, and output that an analyst can review quickly.
  • Delegate with review: tasks with light interpretation, such as suggesting how to classify an entry based on history the accountant has already approved.
  • Keep with the accountant: opinions, guidance on tax treatment, decisions on provisions, signatures, and any response that commits the firm.

Document checks: the agent compares, the accountant decides

Document checking is the most natural case for delegation because it means comparing information against a known list. Invoices, receipts, bank statements, and payroll documents often arrive in mixed formats, and checking each one by hand consumes time that adds no judgment. An agent can do the first sweep and deliver a queue of exceptions, not a list of everything.

A safe setup usually follows these steps:

  • Define, by client type and tax regime, the list of expected documents, written by the firm and reviewed by the responsible accountant.
  • Ask the agent to flag objective discrepancies: a missing field, a date outside the period, a tax ID that differs from the registration record, and an amount that does not match the bank statement.
  • Separate alerts from confirmed errors. Alerts go to the analyst's queue. Confirmed errors enter the process only after human review.
  • Record what was checked, what was found, which version of the rule was applied, and who reviewed it.
  • Do not let the agent correct entries on its own. It suggests, and the professional approves.

These documents contain personal and financial data about clients and third parties. Before sending any file to a model, check where processing happens, who has access to the records, and whether the volume sent is the minimum necessary. The question of transferring data outside the country is covered in international data transfer under LGPD when using AI providers.

Document follow-up: reminders and tracking, not pressure

Much of the delay in receiving documents is not refusal. It is forgetting or unclear instructions about what to send. That is why chasing documents suits an agent: it can send reminders on agreed dates, record who responded, resend the list of what is missing, and alert the team when an internal deadline passes.

The limits need to be just as clear. The agent does not negotiate deadlines, does not threaten, does not promise that a delivery avoids penalties, and does not create deadlines the firm has not set. Active WhatsApp messages require approved templates and prior permission from the contact, and the conversation window has its own rules. For details, see WhatsApp outbound messages: templates, opt-in and best practices.

  • Set a follow-up calendar approved by the responsible accountant, with dates and a maximum number of reminders per pending item.
  • Use only approved wording. If a client asks for something outside the approved text, the agent answers with what is permitted and passes the rest to a person.
  • Treat every reply as a record. A file received goes straight back into the document check.
  • Escalate to a person when there is a dispute, a request for an installment plan, a complaint, or any mention of a fine or audit.

Client communication: the agent answers what is already decided

In client service, the core rule is that the agent answers what the firm has already decided and written down. That includes office hours, the status of a document already received, how to send a file, what information a request needs, and what was agreed in earlier meetings. When a question requires interpreting the client's situation, the right path is a handoff with a summary.

A handoff only works if the person receiving the case does not have to ask everything again. The agent should pass along the client's name, what they asked for, what has already been checked, what information is still missing, and the reason for the handoff. It should also make clear from the first message that the client is talking to an automated assistant, with access to a human when needed.

There is also a specific risk with incoming messages. A client, or someone pretending to be one, can send text designed to override the agent's instructions. OWASP documents this type of attack, called prompt injection, and recommends treating received content as untrusted data. In practice, this means the agent is not allowed to change records, issue documents, or send data without a separate verification step.

Illustrative scenario: a month-end routine at a fictional firm

Illustrative scenario, no real data: a small firm serves service companies with a team of three people. Before the pilot, the team spent much of the month-end week chasing invoices by message, checking files one by one, and answering the same questions repeatedly. The responsible accountant defined three things: the document list for each client type, the follow-up calendar, and the list of topics the agent does not answer.

During the pilot, the agent received files through WhatsApp, checked whether they were legible and complete, and updated each client's status. Pending items followed the approved calendar. Every morning, the analyst received only the exception queue: amount discrepancies, documents dated outside the period, and replies the agent could not classify. One request for an installment plan was passed to the responsible accountant with a summary of the history, and the decision stayed with them.

The expected gain in the pilot was not to eliminate review. It was to focus human attention where it makes a difference. Recording each decision helped the team adjust the document list after two weeks of use, which is the kind of correction that only shows up in real operations.

How to start without exposing the whole portfolio

The path we follow in projects like this is the same for accounting firms and other businesses: map, limit, pilot, and hand over. Mapping means describing the current flow and recording a baseline, such as how much time chasing takes and how many back-and-forth exchanges happen before a document is accepted. Limiting means defining what data the agent can see and what actions it can perform, always starting with the smallest possible set.

Piloting means running a real routine with a defined portfolio, a record of every action, and a success criterion agreed before starting. Handing over means transferring the operation with documentation, named owners, and escalation rules, then deciding whether the routine will be expanded, adjusted, or stopped. This cycle prevents the firm from discovering the agent's limits after the whole portfolio has been exposed.

For governance, NIST publishes a risk management guide for AI that organizes the topic into govern, map, measure, and manage functions. It is not a list of legal obligations, but it helps structure the conversation between the partner, the team, and the responsible accountant about what could go wrong.

Frequently asked questions

Can an AI agent post invoices or do bookkeeping on its own? This is not recommended. The agent can check, compare, and suggest classifications based on approved rules, but the actual posting should go through review by the responsible professional. This separation keeps technical responsibility where it belongs and makes later audits easier.

Can a WhatsApp agent chase clients for documents without authorization? Follow-ups must respect the contact permission and the platform's rules for active messages, using approved templates when the conversation is outside the service window. Before starting, the firm should confirm with each client how and when they will receive messages. The agent also should not promise deadlines or consequences the firm has not set.

Which office tasks are the safest to start with? The safest tasks are usually those with a written rule and an output that is easy to review, such as checking whether the documents on a list are present and legible, and reminding clients of pending items on the approved calendar. Avoid starting with tasks that involve tax interpretation or decisions about a specific client's situation. Those can come later, once the earlier pilot has produced enough records.

How do we start without hiring a large project? Choose one repetitive routine, measure how it works today, define what the agent can see and do, and run a pilot with a small portfolio for an agreed period. At the end, decide whether the routine should be expanded, adjusted, or stopped, based on the records rather than impressions. If the firm wants to understand the terms better before deciding, the superintelligence glossary for executives helps align the internal conversation.

Sources

accounting firmsAI agentssuperintelligenceLGPDworkflow automation