Superintelligence in Local Government: Citizen Service and Limits

Superintelligence in Local Government: Citizen Service and Limits

Superintelligence in local government is, today, a topic for debate, not a product a city hall can buy. What already exists are far more capable AI models that understand informal written requests, consult regulations and organize cases. They can help with citizen service, as long as any decision affecting rights stays with an identified public servant and the use is transparent and compatible with LGPD (Brazil's data protection law).

Key takeaways

• Superintelligence is a future concept. The real short-term gain comes from narrowly scoped AI agents that are well integrated with the municipality's systems.
• In citizen service, AI improves triage, guidance on documents and tracking of case numbers. It does not replace the public servant in decisions about benefits, fines, licenses or taxes.
• Transparency is part of the service: citizens need to know they are talking to an AI, what it can do and how to reach a person.
• LGPD has its own rules for public authorities. Purpose, data minimization, logging and human review must be defined before the pilot.
• The safe path is to map one flow, limit data and actions, run a pilot with an agreed metric, and only then decide whether to expand, adjust or stop.

What does superintelligence mean when we talk about city government?

The term describes a system that would outperform humans at practically every relevant intellectual task. No city government runs anything like that, and no serious vendor should promise it. When someone uses the word in a commercial proposal, ask them to describe, in plain language, which tasks the system performs, with which data and under what supervision.

For a public manager, a different distinction is more useful. Some systems follow fixed rules, like a form that validates fields. Some are language models that interpret free text and generate answers. And some are agents, which combine a model with tools, such as looking up a case number or opening a request. Agents are what actually reaches the city's digital front desk. To understand how these concepts differ, read the difference between AGI, ASI and AI agents, which separates the theoretical debate from what an organization can actually run.

It also helps to stay cautious about expectations. A more capable model writes better, makes fewer mistakes on common tasks and handles ambiguous requests. But it can still be wrong with apparent confidence, invent a rule that does not exist or mix up procedures from different departments. In public service, that kind of error has a cost for the citizen, so greater capability calls for clearer controls, not fewer.

What does more capable AI change in citizen service?

Municipal service often suffers from a familiar problem: citizens do not know which department handles what they need, which channel to use or which documents to bring. Much of the demand is repetitive and depends on information that already exists in portals, decrees and internal manuals, but is scattered. That is where more capable models help.

• Understand the request in everyday language, including long messages, typos or transcribed voice notes, and route it to the right service.
• Guide people on documents, deadlines and steps based on the municipality's official sources, citing where the information comes from.
• Report the status of existing cases by querying the source system instead of answering from memory.
• Gather the required data before handing the case to a public servant, reducing back and forth.
• Serve people outside office hours with basic information and a logged request for the next business day.
• Adapt the wording for people unfamiliar with administrative terms, without changing the content of the rule.

WhatsApp is the most natural channel in Brazil, because citizens already use it every day. That brings channel-specific requirements, such as rules for outbound messages and conversation windows, and the guide on how to integrate AI into your company's WhatsApp covers technical points that apply to the public sector too. Each channel should have an owner, a defined scope and a clear path to a person.

A useful reference comes from a neighboring case: the school office. The same repeated questions about enrollment, documents and the calendar appear there, and the article on AI for schools, admissions and parent communication shows how to separate what the AI informs from what the office decides.

Where is human decision mandatory?

In the public sector, an administrative act has a responsible person, a legal basis and a right of appeal. An AI can support the preparation of a case, but it should not be the invisible author of a decision. The practical rule is simple: the greater the effect on rights, money or sanctions, the greater the human control.

A decision guide to classify each task:

• Inform: explain a service, opening hours, a document or a step from an official source. The AI can answer on its own, with a log and a link to the source.
• Organize: classify a request, summarize a case, suggest the competent department. The AI can do this, with sample review by a public servant.
• Prepare: draft an official letter, opinion or reply. The AI suggests, the public servant reads, edits and signs.
• Decide: grant or deny a benefit, issue a fine, issue a license, define a tax classification. Human decision is mandatory, with the AI at most presenting organized information.
• Act on sensitive data: health, social assistance, children's data. Minimal scope, restricted access and human approval before any action.

LGPD gives people the right to request review of decisions made solely through automated processing that affect their interests. The article on LGPD automated decision-making and the right to review explains what that requires in practice. For a city hall, the most prudent path is to design the flow so that no relevant decision is purely automated, and so that citizens know how to ask for review.

The one who answers to the citizen is still the municipality, not the model.

How do you meet LGPD and transparency requirements in AI-assisted service?

LGPD has a specific chapter on data processing by public authorities, covering public purpose, public interest, information about the situations in which data is processed and the appointment of a data protection officer. The legal basis for each flow should be validated by the municipality's legal team and its data protection officer. The website of the ANPD (Brazil's data protection authority) gathers official guidance that serves as a reference.

A checklist to run before putting any agent in contact with citizens:

• Purpose: which service the agent supports and why the data it asks for is necessary for that service.
• Minimization: the agent only asks for and only sees what the flow requires. Full documents and sensitive data stay out, unless there is a justified need.
• Clear notice: the first message says it is automated service, what it does and how to reach a public servant.
• Official sources: answers based on the municipality's rules and pages, with no improvisation by the model.
• Logging: every conversation and action is recorded, with date, source consulted and referral, for audits and for handling data subject requests.
• Retention: a defined period for keeping conversations and a disposal rule.
• Vendors and data location: a contract with clear roles, and attention to who processes the data and where. If the model runs outside the country, see what LGPD requires in international data transfer when using AI providers.
• Security: role-based access, protection against malicious instructions embedded in messages and periodic review. The OWASP Top 10 for LLM applications lists these risks in an objective way.

Transparency also applies inside the administration. Citizens and oversight bodies should be able to understand which system was used, for what purpose and under whose responsibility. The OECD AI Principles highlight exactly this: transparency, explainability and accountability, and they work well as a basis for an internal policy.

What does it look like in practice? An illustrative scenario

Illustrative scenario: a mid-sized city wants to reduce the volume of repeated questions about reissuing a municipal document. Today, citizens call or go to the counter, and often have to come back because they lack a supporting document.

The agent, on the city's official WhatsApp number, introduces itself as automated service and asks which service the citizen needs. It explains the required documents based on the official service page and sends the link. If the citizen confirms they want to open the request, the agent collects only the name and document number, registers the request in the case system and returns the generated number.

When the person asks something outside the scope, such as disputing an amount charged, the agent does not give an opinion. It summarizes the case, says that a public servant will take over and routes it to the right queue. The public servant receives the summary, opens the case and replies on behalf of the agency. The whole conversation is logged, and the responsible team reviews a sample every week.

In this design, the AI reduces the rework of informing and collecting, but the decision and the signature remain human. The scenario is only an example of architecture, not a result achieved by anyone.

How to start: map, limit, pilot, hand over

The method we apply in AI projects for organizations follows four moves, and it works well in a public agency because it forces decisions before building.

• Map: choose a high-volume, low-risk service. Describe the current flow, who handles it, which systems they use, which questions repeat and how results are measured today. Without that starting point, there is no way to know whether things improved.
• Limit: define which data the agent sees, which actions it can perform, what is forbidden and when it hands over to a person. Close permissions by default and open only what is needed.
• Pilot: run a single real flow, with logs, sample review and a success metric agreed with the department before the start. Include tests with hard questions and attempts to push the agent off course.
• Hand over: deliver the operation with documentation, owners, a review routine and criteria to expand, adjust or stop. The decision to continue belongs to the municipality, based on the pilot records.

To structure risk, the NIST AI Risk Management Framework offers a well-established vocabulary of governance, mapping, measurement and management that fits the routine of an internal committee. When needed, managed operation can be delivered by Be Human, always with restricted permissions, observability and human approval.

Frequently asked questions

Can superintelligence already be used in city governments? There is no evidence that any system at that level exists and is available to buy. What a city can use today are AI models and agents with limited scope, supervised by public servants, to inform, organize and prepare tasks.

Can AI decide a citizen's request, such as a benefit or a license? Decisions that affect rights should have a responsible public servant, a recorded legal basis and a right of appeal. The AI can gather information and suggest a draft, but the analysis and the signature stay with a person. The municipality's legal team should validate each flow.

Do you need to tell citizens they are talking to an AI? Yes, it is the safest practice and consistent with LGPD's transparency principle. The notice should come at the very start, say what the automated service does and indicate how to reach a public servant. It also reduces frustration and complaints.

Where should a city government start? With a high-volume, low-risk service that has clear official sources and a defined owner. Map the current flow, limit data and actions, run a pilot with logs and an agreed metric, and only then decide whether to expand. Sensitive topics, such as health and social assistance, should wait for a later and more controlled stage.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• OECD AI Principles
• OWASP Top 10 for LLM Applications