Superintelligence in Accounting: Month-End Close and Review

Superintelligence in accounting, today, is not a machine that takes over the month-end close. It is increasingly capable AI agents working on organized documents, written routines and per-client memory, preparing and checking what the accountant decides and signs. The gain comes from how well the firm is organized, and technical responsibility stays with the professional.

Key takeaways

• Agents help prepare and review the close, but they do not decide, file or sign.
• Results depend on named documents, predictable folders and a written close routine more than on which model you pick.
• Each client needs its own memory, maintained by humans: tax regime, particularities, recurring pending items and decisions already made.
• The first pilot should be a low-risk review, with a log, an agreed success criterion and mandatory human review.
• Professional secrecy and LGPD (Brazil's data protection law) belong in the design from day one: minimal data, access per client and a record of every action.

What superintelligence means for an accounting firm today

In the strict sense, superintelligence is a system that would outperform people at practically any intellectual task. That is not what a firm operates today. What exists are models and agents that are better at reading documents, comparing information and following long routines without losing the thread. If you want the distinction explained calmly, read the difference between AGI, ASI and AI agents.

For the firm's owner, the useful question is a different one: what does this extra capability change in the close? It changes how much checking can happen before a person looks at the work. An agent can go through every entry in a month, compare them with bank statements and invoices, and list the discrepancies with the source of each. What it does not replace is judgment on classification, tax framing and fiscal risk.

A note on how we read sources in this article. What comes from official documentation is identified as such. The rest is working method and hypothesis that you should validate in your own firm, with your own clients and your own accounting system.

Why start with the close and the review

The monthly close has four traits that make it a good first territory for an agent:

• Repetitive: it happens every month, with similar steps across clients.
• Verifiable: the output can be compared against a source, such as a bank statement, an invoice or the previous month's trial balance.
• Reversible: the error is found before anything goes to the client or to the tax authority.
• Owned: a named person reviews the work and answers for the result.

The opposite is also true. Filing ancillary tax obligations, responding to an audit, issuing an opinion and giving tax advice to a client are not places to start. There, errors are costly, hard to undo and depend on judgment that the firm answers for to the client and to the Regional Accounting Council.

Anthropic, in its material on building effective agents, recommends starting with the simplest solution and adding complexity only when the need is proven. In accounting, that means a narrow review flow, not an agent that tries to do everything.

How to organize documents so the agent can read them

An agent is only as good as the material it receives. In many firms, material arrives as a photo of a receipt sent over WhatsApp, a scanned PDF with a generic name, statements from several banks mixed together and spreadsheets with duplicate versions. An experienced person copes with that. An agent makes mistakes or skips items.

Anthropic treats context as a limited resource and advocates giving the agent the leanest set of information relevant to the task. In other words, an organized folder is not fussiness: it is what lets the agent retrieve only what matters. Practical rules:

• One folder per client and per period, with the same structure for every client.
• A single file naming standard, with client, period, document type and source.
• Statements in exported formats, such as OFX or CSV, whenever the bank offers them, instead of photos or scanned PDFs.
• Documents received by WhatsApp or e-mail saved to the folder with the date received and who sent them.
• A clear separation between what came from the client and what the firm produced, such as reconciliations and supporting spreadsheets.
• Marking of superseded documents, so the agent does not check against an old version.

This organization of documents, memory and permissions is part of what we describe as the company superintelligence layer. It serves the agent, but it also serves the human team, and it usually reduces rework even with no AI at all.

How to build per-client memory

Per-client memory is not the history of a chatbot conversation. It is a living record, written and reviewed by humans, that the agent consults on demand when working on that client. Consulting on demand matters: dumping everything into the context for every task makes the agent slow, expensive and more likely to mix up information.

A useful client record has these blocks:

• Registration data, tax regime and responsible accountant.
• Classification particularities: accounts that always raise questions and the decision already made, with the reason.
• Recurring pending items: the supplier who usually sends invoices late, the partner who pays personal expenses from the company account.
• Client calendar: cutoff dates, who approves, preferred channel, such as WhatsApp or e-mail.
• Correction history: what the agent got wrong or the reviewer adjusted, already turned into a rule.
• Limits: what the agent must not do for that client.

Three precautions keep memory from becoming a problem. First, every item has a date and an author. Second, changes go through review by the responsible accountant, because a wrong rule in the record repeats every month. Third, each client record has an owner, because memory without an owner ages and starts to push errors that look confident.

Step by step: a pre-close with an agent

The plan below is a working hypothesis, not a ready-made recipe. Adjust it to your accounting system and your workflow.

• Step one, collection: the agent compares the documents that arrived for the period against the list expected in the client's memory and produces the list of missing items.
• Step two, reading: it extracts data from invoices, statements and tax payment slips into a standard format and flags what it could not read.
• Step three, cross-checking: it compares statements with entries, invoices with accounts payable and receivable, and balances with the previous month.
• Step four, discrepancies: it lists each discrepancy with its source document, the memory rule that applies and a confidence level.
• Step five, human review: the analyst resolves the discrepancies and the responsible accountant approves or sends back.
• Step six, recording: adjustments become updates to the memory, and the log keeps who decided what.
• Step seven, communication: any message to the client is produced as a draft, and only a human sends it.

Illustrative scenario: a firm serves a dental clinic under Brazil's Simples Nacional regime. During the pre-close, the agent notices a recurring debit on the statement with no matching invoice. The client's memory records that this supplier usually sends the invoice late. Instead of posting it on its own, the agent lists the pending item, cites the note in the record and suggests chasing the supplier through the client's preferred channel. The analyst confirms, the accountant decides whether to keep a provisional entry, and the agent only drafts the message for approval.

Note what this scenario does not contain: the agent did not classify, post or write to the client by itself. It saved the search, brought the right context and left the decision with the person who answers for it.

Where technical responsibility sits and what the agent does not do

Technical responsibility belongs to the registered accountant. No tool changes that, and a firm that tries to shift that weight to the agent is creating risk, not efficiency. A simple decision guide helps the team:

• The agent can do on its own: read, extract, compare, list discrepancies, draft messages and update the pending-items list.
• The agent proposes and a human approves: non-standard classification, adjustment entries and messages to the client.
• The agent does not do: file obligations, sign financial statements, communicate with the tax authority, change registration data or decide tax framing.

Two practices support this design. The first is to measure the agent by sampling: periodically compare its list of discrepancies with what the reviewer found, and record what was missed and what was a false alarm. The second is to never treat the confidence the agent states as proof. It helps prioritize the review, not skip it.

When the output can significantly affect a person, it is worth knowing what LGPD says about reviewing automated decisions. We wrote about it in LGPD automated decision-making and Article 20. The practical principle is to keep a human in the decision and record the criterion used.

Secrecy, LGPD and how to start: map, limit, pilot, transfer

Accounting data carries personal data of partners, employees and your client's customers, on top of the professional secrecy the firm already has to respect. Before switching on any agent, define:

• Access per client, never to the firm's entire archive.
• Minimal data: if the task is reconciliation, the agent does not need the full payroll.
• Where the model runs and whether data is sent outside the country, a topic we cover in international data transfer under LGPD when using AI providers.
• A record of every read and every action, so you can reconstruct what happened.
• An internal AI use policy and clear clauses with the client where appropriate, with legal support.

To organize risk, the NIST AI Risk Management Framework is a voluntary and useful reference for structuring who governs, and how risk is mapped, measured and treated. ANPD, Brazil's data protection authority, is the official source for following data protection guidance in the country.

The method we apply follows four moves:

• Map: describe the current close for a few clients, with steps, owners, tools and the points where rework starts. This is also the baseline for knowing later whether anything improved.
• Limit: define which data the agent reads and which actions it may execute, leaving everything else blocked.
• Pilot: run a single real flow, such as checking the bank statement against entries, with logs and a metric agreed beforehand, such as discrepancies found before human review.
• Transfer: hand the documented flow to the team, with the memory, the logs and the decision criteria, and then choose between expanding, adjusting or stopping.

When a firm prefers not to operate this on its own, managed operation is delivered by Be Human, with the same rules of limits, logging and human approval.

Frequently asked questions

Can AI do the accounting close on its own? That is not the design we recommend. Agents can prepare, cross-check and review, but classification, approval and signature stay with the responsible accountant. A well-built flow reduces search time and makes human review more focused.

Where should a small firm start? With the review of a single document type, for example bank statements against entries, for a few clients whose documentation is already organized. Write the memory record for those clients, define who reviews and agree the success metric before you begin. Only after a few pilot periods is it worth deciding whether to expand.

Do I need to change my accounting system to use agents? Not necessarily. Many pilots work with files exported from the current system, such as reports, OFX and CSV, with no direct integration. Deeper integrations can come later, with restricted permissions and logging of actions.

Does the client need to know the firm uses AI? It depends on the contract, the type of data processed and the legal basis applied, so the right answer involves legal advice. As a practice, we recommend transparency: say that the firm uses support tools, which data goes in and that technical responsibility stays with the accountant.

Sources

• Anthropic: Building effective agents
• Anthropic: Effective context engineering for AI agents
• NIST AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados