
Superintelligence for Accounting Firms: What to Delegate and What to Sign
See which accounting firm routines can go to a layered AI agent, and which signatures and checks stay with the responsible accountant under LGPD.
Want to apply this in your company?
Book a call with Juan →Superintelligence for accounting firms, in practice, is not a system that replaces the accountant. It is a layered architecture in which an AI agent collects, organizes and proposes, fixed rules calculate, and a qualified professional reviews, approves and signs. For the firm, the useful question is which routines can be delegated with logging and clear limits, and which require a signature, a technical decision or professional responsibility that cannot be transferred to a machine.
Key takeaways
- Delegate document collection, suggested classification and follow-up on pending items, as long as everything is logged and stays within defined limits.
- Keep tax and obligation calculations in deterministic rules, never in the free-text answer of a language model.
- Signatures, declarations, opinions and classification decisions stay with the responsible accountant, without exception.
- Treat client data as personal and sensitive business data, under LGPD, with minimum access and a trail of decisions.
- Start with a real workflow, with logs, a success metric agreed in advance, and a clear rule for expanding, adjusting or stopping.
Why the word superintelligence muddies the decision
The term appears often in commercial content and creates the impression that a system exists that can take over the whole firm. That is not how it works today. What exists are language models that are useful for reading, summarizing, classifying and drafting, combined with tools, rules and people. If leadership treats the subject as a staff replacement, implementation tends to start in the wrong place. If it treats it as process design, with clear roles, the gain shows up where it is real.
It helps to separate the language from the operation. To understand what the word can and cannot mean, the article on ChatGPT and superintelligence keeps the conversation at the right scale.
How a layered agent works in the firm
A sensible design separates the work into layers with different responsibilities. The first layer receives and organizes: documents arrive by email, WhatsApp or upload, are identified and linked to the correct client. The second layer proposes: it classifies entries, suggests reconciliations and prepares report drafts. The third layer is deterministic: rates, deadlines and calculations follow validated tables and formulas, with no interpretation by the model. The fourth layer is human: the accountant reviews exceptions, approves entries with impact and makes the decisions. The fifth is the record: every agent action is kept with date, source, rule version and who approved it.
The separation matters because each layer fails in a different way. A model may misread a scanned document. A fixed rule only fails if its table is out of date. A person can tire during a long list. Designing the system so that each error is caught by another layer is more reliable than asking the model to get everything right.
What can be delegated safely
Some routines are repetitive, high in volume, and produce output that can be checked before it becomes an obligation. They are usually good candidates for an agent with limits:
- Receiving and triaging documents, with client identification, document type and date.
- Reminders for pending documents and requests for missing information, respecting contact hours and contact rules.
- Suggested classification of expenses and revenues, always marked as a suggestion until confirmed.
- Suggested reconciliation between bank statements and entries, with differences highlighted for review.
- A summary of each client's obligations calendar, based on data already registered and rules approved by the firm.
- Drafts of client notices and management reports, which the accountant reads before sending.
- First-level answers to frequent questions, with handoff to a person when the topic goes beyond the script.
None of these actions should be allowed to change an entry that is already closed, issue a payment slip or send anything to a government body without approval. Permission to read and draft is different from permission to execute, and the design should reflect that.
What stays with the accountant
Some acts depend on licensing, technical responsibility and signature. These should not be delegated to an agent, however capable it appears:
- Signing financial statements, opinions and declarations that carry professional responsibility.
- Use of the digital certificate and authorization to send filings to government bodies, which must remain under the control of an authorized person.
- Decisions on tax classification, planning and the interpretation of rules that affect the client.
- Approval of adjustments, amendments and closing entries.
- Any communication that makes a commitment, gives a final technical opinion or deals with a conflict with the client.
- Payments and financial transfers, which must depend on explicit human approval.
The practical rule is simple: when the act has legal or tax consequences and requires a professional's name, the agent prepares, but does not conclude.
Illustrative scenario: pending documents at closing
Illustrative scenario: a firm with a portfolio of small businesses receives, close to the closing date, receipts in various formats by WhatsApp and email. Today, a team of assistants spends long stretches identifying which file belongs to whom and chasing clients by message. In a pilot, the agent would triage the files, link each document to a client, list what is missing for each company and prepare follow-up messages for approval. Tax calculations would stay in spreadsheets and rules the firm has already validated. The accountant would review the list of exceptions, such as unreadable documents or a mismatched company tax ID, and decide what to do. Success would be measured by the time until the pending list is closed and by the number of documents linked to the wrong client, two metrics agreed before the start. If the linking error appeared often, the triage layer would be adjusted before any expansion.
Checklist before releasing a routine to the agent
Before putting a routine into production, the firm can answer these questions in writing:
- Is the routine mapped, with inputs, outputs and who approves each step?
- Is the data the agent can access the minimum needed for the task?
- Are the calculations and relevant rules in code or a validated spreadsheet, outside the free text of the model?
- Is there a list of forbidden actions and a list of actions that require human approval?
- Is each action logged with date, source, data used and the person responsible for review?
- Is there a simple way to switch the agent off or return to the manual process?
- Does a person at the firm know the exceptions and what to do with them?
- Were the success metric and the stopping criterion agreed before the pilot?
If any answer is no, the routine is not ready for delegation, however promising it looks.
Client data, LGPD and recording decisions
Accounting firms handle personal data, company data and sensitive tax information. LGPD applies to the processing of this data, and the Brazilian National Data Protection Authority, the ANPD, is the reference channel for guidance. In practice, this means defining the purpose, limiting the agent's access to what is necessary, checking where the data is processed and keeping a trail of decisions. When an automated decision affects a client, it is worth understanding what Article 20 of LGPD provides on review.
For managing risk in AI systems, the NIST AI Risk Management Framework offers a well-known structure for identifying, measuring and treating risk across the system lifecycle. It is not a legal checklist, but it helps organize the conversation between leadership, the responsible accountant and whoever operates the tool.
In how we work, the sequence is to map the current process and its baseline, limit data and actions, pilot a real workflow with logs and an agreed metric, then hand over the operation and decide whether the workflow should be expanded, adjusted or stopped. This logic applies to accounting as much as to any other sector. It also prevents a pilot from becoming a permanent exception with no owner.
For those just starting, the overview of superintelligence in accounting firms details how to prepare the tax routine and the client relationship. Standardizing data beforehand also makes a difference, and Anthropic's engineering write-up on building effective agents discusses design principles for multi-layer systems that are worth reading.
Frequently asked questions
Does superintelligence in accounting already replace the accountant? No. What exists today are AI agents that help with collection, triage, suggested classification and drafts, always combined with fixed rules and human review. Professional responsibility for signatures, declarations and opinions remains with the licensed accountant, and no serious design should take that away from them.
Can the agent calculate taxes and generate payment slips? The calculation must come from deterministic rules and validated tables, not from the free-text answer of a language model. The agent can prepare the check, organize the input data and flag discrepancies. Issuing and sending payment slips depends on explicit human approval and on logging the action.
How do you ensure LGPD compliance when using AI with client data? Start by defining the purpose of the processing and limiting the agent's access to the data strictly needed. Check where the data is processed, record decisions and keep a channel for review whenever there is a relevant effect on the client. Official guidance is available on the ANPD website, and the firm's specific legal analysis should be done by a qualified professional.
Where should you start without promising results? Choose a repetitive, low-risk routine, such as document triage, map the current process and measure time and errors before changing anything. Pilot with logs, a metric agreed in advance and a designated owner. At the end of the period, decide whether the routine should be expanded, adjusted or stopped, based on the records and not on a general impression.