Superintelligence for Law and Accounting Firms: What AI Absorbs

Superintelligence for Law and Accounting Firms: What AI Absorbs

Superintelligence for law and accounting firms is, today, a working hypothesis, not a product you can buy. The lens is still useful because it forces a practical question: if AI keeps getting more capable, which parts of your routine does it absorb, and which parts still depend on a licensed professional who answers for the result? In general, reading, checking, drafting and organizing tend to be absorbed, while opinions, signatures and technical responsibility stay with a person.

Key takeaways

• No current system is superintelligent. Using the idea as a planning scenario is valid, as long as you keep clear what is documentation, what is implementation experience and what is hypothesis.
• Reading, triage, checking and first drafts are the parts of the routine most likely to be absorbed by AI, with or without superintelligence.
• Technical responsibility, professional secrecy and the trust relationship with the client do not transfer to a model. Whoever signs is accountable.
• The central risk moves: it leaves the effort of producing and lands on the ability to review well. Reviewing without time or criteria becomes the weak point.
• The safe path is to map the process, limit data and actions, pilot one real workflow with a record of what happens, and only then decide whether to expand, adjust or stop.

What does superintelligence mean for a regulated profession?

The term describes a system that would outperform humans at most relevant intellectual tasks. That does not exist yet as a verifiable product, and any vendor selling it as a current deliverable deserves suspicion. To understand the difference between general intelligence, superintelligence and the agents a company actually operates, read the article on the difference between AGI, ASI and AI agents.

For a firm, the useful question is not when superintelligence arrives, but which organizational decision you would make today that would still make sense if model capability doubled. That is the lens. It separates three layers:

• Official documentation: what your profession's regulations, the ANPD (Brazil's data protection authority) and model providers state in writing.
• Implementation experience: what you observe when a real workflow runs, with logs, errors and review.
• Hypothesis: what may happen if capability keeps improving, which should guide preparation, not promises.

Mixing the layers is the most common mistake. An optimistic article about the future does not replace reading the rules of the Brazilian Bar Association (OAB) or the Federal Accounting Council on how you practice.

What increasingly capable AI tends to absorb from the routine

The pattern is consistent: tasks where the input is text or structured data, quality can be checked and errors can be caught in review are the first to move.

In law, this includes:

• Reading and summarizing case files, contracts and documents sent by the client.
• Triage of court publications and case updates, classified by urgency and deadline.
• First drafts of filings, notices and contracts based on the firm's own templates.
• Internal search across filings and opinions the firm has already produced.
• Document organization and checklists by type of matter.

In accounting, the list is usually:

• Classification and reconciliation of entries from bank statements and invoices.
• Checking tax documents against known rules and inconsistencies.
• Chasing clients for missing documents by email or WhatsApp.
• Drafting replies to recurring questions about routines, due dates and obligations.
• Building management reports from data that has already been validated.

None of this removes the need for checking. It means the professional's time shifts from producing the first draft to verifying, correcting and deciding.

What still requires human technical responsibility?

Some acts exist in a regulated profession precisely so that an identifiable person answers for them. Even with much better models, the question of who is accountable does not go away.

• Legal or accounting opinions and advice to the client, which involve interpretation, risk and context.
• Signing and filing of pleadings, financial statements, tax returns and other acts that require a license.
• Strategy: choosing between settlement and litigation, or between one tax regime and another, based on how much risk the client tolerates.
• Professional secrecy: deciding what can leave the firm and where it can go.
• Trust: delivering bad news, negotiating, declining a matter for ethical reasons.
• Judgment on exceptions, when the case departs from the pattern the model has seen.

If you want to go deeper on the legal side, the article on Brazilian Bar Association AI guidelines for lawyers covers what lawyers should observe. For accounting, the equivalent is to follow the guidance of your professional council and the requirements of your own practice.

One practical point: a more capable AI can also be wrong with more conviction. Nonexistent citations, outdated rules and plausible but wrong numbers remain failure modes. The more fluent the text, the more disciplined the review needs to be.

How do you decide what to hand to AI and what to keep with the team?

A simple decision guide, applied task by task, avoids abstract debates. For each task in the firm, answer in order:

• Is a reserved professional act involved, such as signing, filing or giving an opinion? If yes, AI only prepares, and the professional decides and signs.
• Would an error be caught in review? If yes, the task is a candidate for automation with mandatory review.
• Would an error cause irreversible harm, such as a missed deadline or a return already submitted? If yes, it requires human approval before the action.
• Does the task use personal or confidential data? If yes, define beforehand where the data is processed and under what legal basis.
• Is there a quality criterion you can measure? If not, it is not yet time to automate.
• Does the decision affect a person's rights in an automated way? Then it is worth knowing the right to review under LGPD (Brazil's data protection law), explained in LGPD automated decision-making and the right to review.

The result is usually three groups: tasks AI performs and the team samples, tasks AI prepares and the team approves, and tasks that stay fully human.

Illustrative scenario: chasing documents in an accounting firm

Illustrative scenario, not tied to any real client. A small accounting firm loses time every week asking clients for invoices, statements and receipts, and clients reply late. Today, one person on the team sends manual messages and checks what arrived.

A pilot workflow could run like this: the system lists, per client, the documents still missing for the monthly close; an agent drafts the request in the firm's own tone; clients who have given permission receive it on WhatsApp, and the rest by email; incoming files are checked against rules (format, period, legibility) and forwarded to the responsible person. The agent does not finalize entry classifications, does not answer questions about tax interpretation and, for any client question outside the script, hands off to a human with the history attached.

What this design preserves: the accountant remains responsible for the calculation and the filing. What changes: the person stops spending the day chasing and starts reviewing exceptions. For outbound messages, the firm needs to handle permission and message templates, the subject of the article on WhatsApp outbound messages, templates and opt-in.

Secrecy, LGPD and where client data ends up

Law and accounting firms handle personal, financial and sometimes sensitive data. Before any productivity gain, three questions need written answers:

• Which tool processes the data, and under what contractual terms? Free personal accounts are rarely suitable for client material.
• Does the data leave Brazil? If so, there are implications covered in the article on international data transfer under LGPD when using AI providers.
• Who on the team may use what? A short internal policy, with what is allowed and what is banned, prevents the improvised use that is behind most incidents.

For official guidance on data protection, the ANPD website is the reference. To structure risk management for AI systems, the NIST AI Risk Management Framework and the OECD principles are recognized starting points, although they do not replace legal analysis of your own case.

The most capable model in the world does not solve a governance problem. If the firm does not know which documents were sent to which tool, better technology only increases exposure.

How to apply the method: map, limit, pilot, hand over

The approach we use in AI projects for companies fits regulated professions well, because it starts from control, not from promises.

• Map: describe the current process, who performs each step, where rework happens and what the baseline is (time spent, volume, types of error). Without this, there is no way to know whether anything improved.
• Limit: define which data the system can access, which actions it can take and which require human approval. In firms, this includes drawing a clear line between preparation and professional acts.
• Pilot: run a single real workflow, with a record of the agent's decisions, sample-based review and a success criterion agreed before starting.
• Hand over: deliver the documented workflow to the team, with named owners, and decide explicitly whether to expand, adjust or end it.

For an internal use policy you can build on, the article on company policy for employee ChatGPT use helps separate what to allow from what to ban. Once the agent is in continuous operation, monitoring and maintenance are part of the work, not an extra.

Frequently asked questions

Will superintelligence replace lawyers and accountants? There is no basis to claim that, because superintelligence is still a hypothesis and regulated professions have acts that require a license and personal responsibility. The more likely outcome is that the mix of work changes, with fewer first drafts produced by people and more review, decision-making and client relationship. The firms best prepared are the ones that organize processes and review criteria now.

Where should a small firm start? With the lowest-risk, highest-repetition workflow, such as chasing documents, triaging incoming requests or summarizing received material. Map the current process, define what the AI must not do and run a short pilot with a record of what happens and human review. Only after measuring the result is it worth expanding.

Can I put client documents into any AI tool? Not before checking the contract, where the data is processed, the legal basis and your profession's secrecy rules. Client data calls for tools with suitable contractual terms and a clear internal policy on who uses what. When in doubt, consult the guidance of the Bar Association, the accounting council and the ANPD.

How do I stop the team from trusting AI answers too much? Make review mandatory for everything that leaves the firm, require checking of any cited sources and rules, and log the errors you find so you can adjust the workflow. Treat model output like a draft from a new colleague: useful, fast and subject to checking before any signature.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• OECD AI Principles