Superintelligence in Accounting Firms: Prepare Tax Routines and Clients

Superintelligence in accounting firms is still a discussion about the future, not a product you buy. What is worth preparing today is the foundation that lets you absorb AI models that keep getting more capable, safely: mapped processes, client memory kept in order, clear limits on what the AI may do, and human review wherever tax responsibility is involved. That foundation serves today's AI as well as whatever comes next.

Key takeaways

• Superintelligence is not a tool to hire tomorrow. The preparation that matters is the same for more capable models.
• Map the tax routine step by step before putting AI into any of them.
• Organize each client's memory with what the work requires, and nothing beyond that.
• Let the AI read and suggest, while people keep approval for sending, filing and any change in tax systems.
• Test on a real workflow with a record of decisions, then decide whether to expand, adjust or stop.

What changes when AI becomes more capable in the office

More capable models can read longer documents, cross information from more sources and follow more complex instructions. That sounds like good news for accounting, which runs on checking, classifying and chasing missing items. The problem is that capability does not replace control. A model that makes a mistake with high confidence in an entry or in the classification of an invoice does more damage than one whose errors are obvious.

For that reason, the useful question for the owner of an accounting firm is not when superintelligence arrives. It is which parts of the routine can already be described clearly, which data is used, who makes the decision, and how you know the decision was good. Firms that can answer those questions today can switch models, widen the use of AI or narrow its scope without rebuilding everything. To understand the difference between what exists today and what is speculation, read ChatGPT is superintelligence? What it is and what it is not.

Map the tax routine step by step before automating

The routine of an accounting firm is often described in general terms, such as "closing the month" or "handling the tax work." That is not enough to decide where AI helps. The mapping has to separate each step, with its input, output, owner and internal deadline. Without it, automation only speeds up what was already confusing.

A good starting point is a checklist like this:

• Document intake: where documents come from (email, WhatsApp, the client portal, a shared folder) and who confirms they are complete.
• Invoice classification and entries: which rules are already fixed and which depend on judgment.
• Bank reconciliation: which differences recur and which require investigation.
• Tax calculations and recurring filings: which calculations are deterministic and which have exceptions that need review.
• Chasing missing items: who is reminded, through which channel, how often and in what tone.
• Client support: which questions repeat and which need the accountant.

For each step, note what is a fixed rule and what is judgment. A fixed rule can be handled by a deterministic workflow, with the AI only helping to read the document. Judgment needs human review and a record of why the decision was made.

Client memory: what to record and what not to keep

Continuity in accounting depends on memory: tax regime, the client's particularities, open items, decisions made in meetings, and exceptions that were already approved. Today that memory is usually scattered across emails, spreadsheets, notes and the mind of whoever handled the account. AI models work better when they receive a short, current and well-organized context, rather than a huge file.

The care required here is both technical and legal. Record what changes the way you work with that client, and avoid accumulating personal data without a purpose. Under Brazil's General Data Protection Law (LGPD), processing must have a clear purpose and be limited to what is necessary. The national data protection authority (ANPD) publishes guidance worth following, and the right to review automated decisions is a topic every firm should understand before letting an AI recommend anything that affects a client.

A practical rule: for each item in the memory, ask which step of the routine it serves and who needs to see it. If the answer is vague, the item probably should not be there.

Limits: what the AI reads, what it suggests and what it never does alone

Most of the risk in AI automation comes from permissions that are too broad. If the agent can read every folder, message every client and change entries, any error spreads. The OWASP Foundation maintains a list of risks for applications built on language models, including prompt injection and leakage of sensitive data, which anyone designing this kind of workflow should read.

In practice, a firm can separate three levels:

• Reading: the AI accesses specific documents of the client under review, without browsing other folders.
• Suggesting: the AI proposes a classification, a collection message or a summary, always marked as a draft.
• Restricted action: sending, transmitting, changing a system or signing stay out of the AI's reach, or require explicit approval from a person.

Illustrative scenario: an accounting firm with a portfolio of small businesses lets the AI summarize bank statements and flag discrepancies in a working spreadsheet. The AI has no access to the bookkeeping system or to the client's digital certificate. The accountant reviews the spreadsheet and only then posts the entries. If the model makes a mistake, the error stays in the draft, where it is cheap to fix.

Human review: who approves what

Human review should not be a rubber stamp at the end of the process. It needs criteria. A simple decision guide helps the team know when to stop and when to continue:

• If the task involves a tax return, a payment or a submission to a government body, the responsible person approves it first.
• If the AI recommends anything that changes the client's tax classification, review is mandatory and the reason is recorded.
• If a reply to the client concerns a deadline, an amount or a penalty, an accountant reviews the text.
• If the AI cannot find the document or the data is incomplete, the case goes to a human queue, with no filling in by assumption.
• If the same correction keeps appearing, the rule or the prompt should be revised, not just corrected case by case.

The decision record is what makes improvement possible. Note what the AI suggested, what the person did and why. That trail also protects the firm if a client questions an entry.

How to set up a pilot in the tax routine

A well-designed pilot has a small scope, a metric agreed before it starts and a date to decide. The sequence below works as a roadmap:

• Choose a single step with recurring volume and relatively clear rules, such as triaging incoming documents.
• Define what success means before starting, for example, the share of documents correctly identified in human review, measured on a weekly sample.
• Limit the AI's access to the documents of the chosen step and to the participating clients.
• Record every suggestion, every correction and every case passed to a person.
• Review the records at fixed intervals with the people who do the task, not only with whoever bought the tool.
• At the end of the period, decide: expand to another step, adjust the workflow or stop.

This is the method we use in AI projects for companies: map the current process and its baseline, limit data and actions, pilot a real workflow with records and an agreed metric, then hand over the operation and decide between expanding, adjusting or stopping. The goal is for the decision to rest on evidence from the process itself, not on generic demonstrations. If you want to see how this work is organized in practice, the page on how we work with companies describes the steps.

Client communication and personal data

Much of the contact between a firm and its clients happens on WhatsApp, and that is where AI tends to show up first: document reminders, answers about deadlines, confirmation of receipt. That channel requires care with platform rules. Messages sent outside a conversation the client started are restricted, as explained in the guide on the 24-hour window on the WhatsApp API.

It is also worth defining what the agent never answers: tax amounts, interpretation of legislation for a specific case, conflict situations, or any request involving sensitive data. In those cases, the conversation should be handed to a named person with a clear return time. Clients notice a clear handoff faster than an evasive answer.

Frequently asked questions

Can superintelligence already be used in an accounting firm? In practice, what firms use today are language models and AI tools, which are useful for reading documents, summaries and drafts. Preparing processes, memory and human review serves any evolution of these models, so it does not depend on a specific future technology. Start with what already exists and can be measured.

Do I have to replace my accounting system to use AI? Not necessarily. Many firms begin with AI reading documents and generating drafts in working files, while entries stay in the current system. Direct integration can come later, if the pilot shows value and the permissions are well defined. Before any integration, confirm how the data will be handled.

Can AI decide a client's tax classification? It should not decide alone. Tax classification has a direct impact on the client and requires professional responsibility. AI can suggest and flag discrepancies, but the decision and the record of its reason should stay with a qualified person. The right to review automated decisions reinforces this logic.

How do I know whether the pilot worked? Agree on the metric before starting, based on the team's real work, such as the correction rate in human review or the time to finish a step. Track the records over a defined period and compare them with the baseline from before. If the result does not support expansion, narrow the scope or close the pilot without embarrassment.

Sources

• Anthropic: Effective context engineering for AI agents
• OWASP Top 10 for LLM Applications
• NIST AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados