Superintelligence in Notary Offices

Superintelligence in notary offices does not mean a system that decides on its own. It means using increasingly capable AI to organize protocols, track deadlines and report the status of requests, while the legal review, drafting and signing of acts that require public faith (fé pública) stay with the responsible notary or registrar. The line is drawn by function, not by technology.

This article uses the Brazilian model, where notary and registry offices (cartórios) operate under a public delegation and local oversight bodies called corregedorias. If you work under a different system, the principle still holds: map what the AI may touch, and keep legally binding acts with the person who is authorized to perform them.

Key takeaways

• AI fits well in the intake and follow-up layer: receiving requests, checking documents, reporting status and warning about deadlines.
• Legal review, the decision on the act and public faith stay with the delegated officer and the authorized staff. AI can at most support reading.
• The protocol must be a structured record with a number, date, owner and status, so that people and AI refer to the same request.
• Legal and regulatory deadlines come from the law and the local corregedoria. AI reminds and flags, but does not define or interpret them on its own.
• The safe path is to map, limit, pilot one flow and only then expand, with a log of every action and human review at the risk points.

What does superintelligence mean at a notary office counter?

In the literature on the topic, superintelligence describes a system that would outperform humans at most cognitive tasks. That is a hypothesis, not a product you can buy. What exists and is within reach of a notary office are language models and agents that can read documents, summarize, classify, draft messages and call tools connected to the internal system.

For the head of an office, this distinction matters for a practical reason. If the conversation starts with an intelligence that solves everything, the project begins without limits. If it starts with specific tasks that have an input, an output and an owner, it can be tested. To understand the difference better, read Is ChatGPT Superintelligence? What It Is and What It Is Not and Superintelligence Myths for Business Leaders.

The starting point, then, is to treat AI as a layer that organizes work. It does not replace the delegation granted by the public authority, and no tool has public faith. The delegated officer has it, and it is exercised through acts performed by the people the law authorizes.

Which steps of a request can AI organize?

A request at a notary or registry office usually goes through recognizable steps, whatever the specialty (notarial deeds, civil registry, real estate registry, protest of debts, titles and documents). The citizen asks a question, submits documents, the request is logged, it goes through checking and legal review, the act is performed or the document is issued, and the result is delivered. The exact rules vary by specialty and by state, so confirm them with the corregedoria and the office's legal counsel.

In this flow, AI tends to be useful in the support steps:

• Answering frequent questions about required documents, channels and how to submit a request, based on content reviewed by the office.
• Checking that received attachments are legible and complete against a previously approved list of requirements.
• Opening the request record in the system by filling in structured fields, for human review.
• Telling the requester where things stand, using only the status recorded in the system.
• Alerting the team about stalled requests, unanswered requirements and deadlines close to the limit.
• Preparing draft messages, such as a request for a missing document, always reviewed before sending when there is a legal implication.

The most natural channel for the service side is usually WhatsApp, where the citizen already is. The guide on How to Integrate AI into Your Company's WhatsApp covers the technical design, and the logic applies to a notary office: the agent informs and organizes, and hands over to a person when the conversation goes beyond what was authorized.

What must stay with public faith?

Public faith is the presumption of truthfulness attached to the acts of the delegated officer. It depends on who performs the act, how it is performed and who answers for it. So the design rule is direct: AI does not perform, sign or attest.

In practice, the following stay outside the autonomy of AI:

• The legal review of the deed or request, meaning the decision on whether it can or cannot be performed.
• Identifying the parties and verifying their capacity and intent in acts that require presence or in-person checking.
• Drafting, signing, certifying and issuing certificates that carry the value of an act.
• Decisions on formal requirements, refusals and doubts, which have consequences for the citizen.
• The interpretation of legal deadlines, priority-filing deadlines or any deadline that produces a legal effect, whenever there is doubt.

AI can help the team read a long document, point out passages that look inconsistent or compare it with a checklist. But the output is a suggestion marked as such, and a qualified person decides. If the office wants to go further, the first step is to consult the corregedoria and legal counsel, and only then design the flow.

There is also the matter of automated decisions. LGPD, Brazil's data protection law, gives people the right to request review of decisions made solely through automated processing that affect their interests. LGPD Automated Decision-Making: The Right to Review (Art. 20) explains what this requires of a flow. In a notary office, the safest path is to leave no decision that affects a citizen exclusively in the hands of the system.

How do you structure protocols and deadlines so AI can track them?

AI can only track well what is recorded consistently. If progress lives in parallel spreadsheets, loose messages and the memory of whoever is at the counter, the agent will repeat the mess, only faster. The groundwork is to give each request a clear identity and life cycle.

A minimum protocol record includes:

• Protocol number and date and time of entry.
• Type of service and the responsible specialty.
• Requester and authorized contact channel.
• Current status, chosen from a closed list (received, under checking, with requirement, awaiting reply, under review, completed, delivered).
• Current human owner.
• Applicable deadline and the source of that deadline (regulation, internal rule or service commitment).
• Action history, showing who did what, when and through which channel, including when it was the agent.

Two decisions make a difference. First, separate the regulatory deadline from the internal one. The first comes from the law or the corregedoria and is entered by the responsible team. The second is the office's own service target. AI can alert on both, but only the first carries legal weight, and a person defines it. Second, use closed statuses instead of free text, so that every answer to the citizen is a reading of the system and not an inference by the model.

The agent also needs narrow permissions: check status, draft a requirement notice, log a note, open a task for the team. It should not be able to change a deadline, mark a request as completed or issue a document. OWASP's guidance on authorization describes the principle of least privilege, which applies directly to these tools.

Illustrative scenario: a certificate request with a missing requirement

What follows is an illustrative scenario, not a real case. A person writes to the office on WhatsApp asking for a certificate and sends a photo of an ID document that is cropped and blurry.

The agent identifies the type of service from the text, checks the attachment against the approved list of requirements and sees that the image is not legible. It replies asking for a new photo and explains why in plain language. When the proper image arrives, it opens the protocol with the structured fields and places the request in the checking queue with the matching status.

A clerk checks the data, decides on the request and records the result in the system. Days later, the requester asks how it is going. The agent reads the recorded status and reports exactly that, without estimating anything not in the system. If the request is close to the limit of a registered deadline, the agent has already alerted the team earlier through an internal task.

At no point in this flow did AI review a deed, issue the certificate or vouch for anything on behalf of the office. If the requester insists on a legal question, the agent passes the conversation to an attendant with a short summary of the history, so nobody has to repeat everything.

What data and security precautions does a notary office need?

Notary and registry offices handle sensitive personal data and documents that stay in circulation for a long time. Before any pilot, the office needs to know which data the agent will see, where it is processed, how long it is retained and who is responsible for the processing. The Brazilian data protection authority (ANPD) publishes official guidance on LGPD, and the office's legal counsel should confirm the legal basis and the sector's specific obligations, which may include rules from the corregedorias.

A few practical points:

• Collect the minimum: a service agent does not need the full content of a deed to report its status.
• Define where processing happens. If data goes to providers outside the country, read International Data Transfer Under LGPD When Using AI Providers. Models running on your own infrastructure are an alternative to evaluate.
• Treat every received document as untrusted input. An attachment may contain hidden instructions that try to manipulate the model, a known risk called prompt injection, described by OWASP.
• Log every query and action of the agent, without exposing unnecessary data in the logs themselves.
• Guarantee the right to speak to a person at any point in the conversation.

The NIST AI Risk Management Framework is a useful reference for organizing this internal discussion: map risks, measure them, treat them and monitor them. It is not a mandatory standard for Brazilian notary offices, but it helps structure governance.

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

The method we use in consulting fits a notary office well, without any promise of results.

Map: describe the current flow for one type of request, who does what, in which system and within what deadline. Set a baseline before any AI, for example how many status questions arrive per day and how much time the team spends answering them.

Limit: decide which data and actions are allowed. Each tool the agent has is a permission. The list of acts reserved for public faith goes in writing as a prohibition, and the agent is designed to route, not improvise. Anthropic's article on building effective agents recommends starting with simple solutions and adding autonomy only when needed, which fits this context.

Pilot: choose a single flow, such as reporting the status of certificate requests, with logs and a metric agreed in advance. Review the cases where the agent handed over to a human and the ones where it got something wrong. The general supporting architecture is described in Company Superintelligence Layer Architecture: Five Building Blocks.

Hand over: give the team the documentation, access, review criteria and monitoring routine, and decide explicitly whether to expand, adjust or stop. Stopping is also a legitimate result of a pilot.

Frequently asked questions

Can AI replace the clerk or the notary? Not in acts that require public faith. AI can support organization, preliminary checking and communication, but the legal review and the performance of the act remain with the people the law authorizes. The flow design should make that explicit.

Can the agent tell citizens about deadlines? It can report what is in the system, such as the request status and the date entered by a responsible person. It should not estimate legal deadlines on its own or interpret regulations. When in doubt, the right move is to hand over to an attendant.

Where should a notary office start? With a single low-risk flow, such as information on required documents or checking request status, with a structured protocol and an action log. Before that, align with legal counsel and check the rules of the local corregedoria. After the pilot, decide with data whether to expand.

Do you need a local model, or is a cloud service enough? It depends on the type of data, the legal assessment and the office's technical capacity. Local models reduce data exposure but require your own operation, as explained in Running Ollama Local Models in a Company: A Practical Guide. The choice should come after mapping which data the agent really needs to see.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• OWASP: Authorization Cheat Sheet
• Anthropic: Building effective agents