Superintelligence for City Governments: What a Citizen Agent Should Do

In city government, an AI agent is well suited to guide residents on services, required documents, service channels, and deadlines that are already published, always citing the official source. It should not decide on benefits, fines, appeals, exceptions, or the interpretation of rules. Those topics must be routed to a civil servant with the context already organized. This division only holds up with a written policy and a record of every conversation, so the administration can explain what was said and why.

Key takeaways

• The agent answers from official, published content and names the source of each piece of information.
• Administrative decisions, exceptions, appeals, and cases involving sensitive data go to a civil servant, without exception.
• Every interaction creates a record: question, answer, source used, handoff, and reason.
• Residents know they are talking to an AI and how to reach a person.
• Before expanding the use, it makes sense to start with one routine, agreed metrics, and periodic review.

Is superintelligence the right term for what the city will use?

In public debate, superintelligence usually refers to hypothetical systems far beyond human capability in any task. No public agency needs that to serve residents. What exists today are language models combined with rules, knowledge bases, and record systems, and that is what needs to be evaluated. If the term appears in planning documents, translate it into operational questions: what the agent answers, what it does not answer, who reviews it, and where the history is stored. For a more careful reading of these terms, the article on superintelligence myths helps separate the conceptual debate from the practical decision.

What the agent can guide residents on

The agent is useful when it answers questions that have a published and verifiable answer. In general, this includes:

• requirements and documents listed on the official page of each service;
• office hours, addresses, and service channels of the agency;
• step-by-step instructions to book an in-person appointment or open a case file;
• status checks on a case file, when an authorized integration exists;
• guidance on where to find official forms and fees already published by the agency itself.

One point needs attention: the agent should answer from a content base approved by the responsible team and marked with a last-updated date. If the official page changes and the base does not, the answer becomes wrong while still sounding certain. For that reason, the base needs an owner and a review routine.

What must be routed to a civil servant

The practical rule is simple: when the answer depends on interpretation, judgment about a specific case, or decision-making power, the agent does not respond as if it were the decision. It collects the essentials and routes the case. Examples on this list include:

• requests for review, appeals, and challenges to fines or charges;
• requests for exemptions, benefits, or exceptions that depend on discretionary criteria;
• questions about how a rule applies to a particular situation;
• reports of risk to a person, such as violence or an urgent health situation, which must follow the agency's emergency protocol;
• formal complaints and requests that involve data about third parties.

Brazil's LGPD grants the right to request review of decisions made solely on the basis of automated processing that affect the data subject's interests. For that reason, the agent cannot be the final word in any of these cases. The topic has its own implications, which the article on automated decisions and Article 20 of the LGPD covers in detail.

How to record each interaction for transparency

In this context, transparency means the administration can reconstruct an interaction later, without relying on anyone's memory. A good record does not keep everything. It keeps what is necessary. A general reference for organizing this kind of control is the NIST AI Risk Management Framework, which structures AI risk management around governing, mapping, measuring, and managing. In practice, the minimum record for each interaction should contain:

• date, time, and channel of the interaction;
• the resident's question and the answer given by the agent;
• the official source used in the answer and the version of the base in force;
• the reason for the handoff and the department that received the case;
• whether the resident asked for a human and whether one was actually offered;
• the summary passed to the civil servant, so nobody has to make the resident repeat everything.

Personal data should appear in the record only when needed, and identifiers such as tax numbers can be masked. The people who can access the logs also need to be defined. A record nobody reads is not transparency, and a record everyone reads without criteria is a privacy problem.

Illustrative scenario: a second copy and an exemption request

Illustrative scenario: a resident contacts the city through WhatsApp and asks for a second copy of a payment slip. Then they ask whether they qualify for an exemption because of low income. The agent explains, based on the official page, how to get the second copy and which channels are available. On the exemption, it explains that the responsible team performs the assessment, lists the documents published as required, and does not say whether the request will be approved. It then opens a handoff with a summary: service requested, documents the resident has already provided, the exemption request, and the preferred return channel. The civil servant responds within the published timeframe, and the record shows each step, including the source used and the reason for the handoff.

What technical limits stop the agent from overstepping?

An instruction written in the prompt, such as telling the agent to never grant exemptions, is not a reliable barrier. Messages from residents may, intentionally or not, try to persuade the agent to ignore those instructions, and OWASP documents this class of attack as prompt injection. The protection has to live in the architecture. The agent should have read-only access to the official base, should not be allowed to change registrations or approve requests, and every sensitive action should pass through a human approval step. This rule applies even when the conversation seems convincing.

How to apply the method in a pilot

The method we use follows four steps: map, limit, pilot, and transfer. First, map the current service, with approximate volume by type of request and the bottlenecks residents face. Next, define the data and actions the agent may use. In the pilot, choose a single service, with logs and an agreed metric set before launch, such as the time to a correct handoff. At the end, the team decides whether to expand, adjust, or stop. This path is slower than a broad launch, but it lets you explain each decision to the department head and the internal audit office.

Frequently asked questions

Can the agent answer questions about taxes and charges? It can answer what is published, such as deadlines, payment locations, and required documents, always citing the source. Calculating the amount owed, payment plans, or disputes should go to the responsible department, with the case already recorded.

Do residents need to know they are talking to an artificial intelligence? Yes. This is the recommended practice and the one that gives the administration the most security. The opening message should say it is an automated assistant, explain what it does, and show how to reach a civil servant.

What data should be kept in the interaction history? Keep what allows the handoff decision to be reconstructed: question, answer, source, reason, and department. Personal data should be limited to what is necessary and masked whenever possible, with access restricted to people who have a role in the case.

How do you start without building a large project? Choose a high-volume, low-risk service, such as guidance on documents, and run a pilot with logs and an agreed metric. The pilot will show what the official base needs to improve before any expansion. An article on AI customer support and what to automate helps separate the questions an agent can answer from those that require a person.

Sources

• NIST AI Risk Management Framework
• OWASP: Prompt Injection Prevention Cheat Sheet
• ANPD: Autoridade Nacional de Proteção de Dados