Superintelligence in Public Agencies: Limits, Records and Human Review

Superintelligence in Public Agencies: Limits, Records and Human Review

How public agencies can use AI for citizen services with clear limits, logged interactions, transparency and human review of decisions.

•

Want to apply this in your company?

Book a call with Juan →

Superintelligence, in the sense of an AI that surpasses human capability and decides on its own, is not a tool a public agency can buy. What can be used responsibly are narrow AI systems limited to service tasks such as triaging requests, explaining procedures and routing cases to the right department. Safe use depends on explicit limits on what the AI may answer, a record of every interaction, a clear notice to the citizen, and human review of any decision that affects rights.

Key takeaways

  • Superintelligence is a reference concept, not an off-the-shelf product. In practice, an agency works with narrow AI, configured for defined tasks.
  • AI can inform, triage and route, but it should not grant, deny or sanction anything without review by a responsible person.
  • Every automated interaction needs a record: what was asked, what was answered, which official source supported the answer, and who took over the case.
  • Transparency starts with the notice to the citizen: they need to know they are talking to an automated system and how to reach a staff member.
  • A pilot on a single service, with a metric agreed in advance and a defined stop rule, is safer than a broad launch.

Why the word superintelligence gets in the way of decisions

In the press and in sales presentations, the word is used to describe anything that sounds advanced, which confuses the people who must buy, authorize or oversee these systems. A director who hears superintelligence tends to imagine a system that thinks on behalf of the agency, and that leads to wrong expectations about autonomy, maintenance effort and accountability. A more useful question replaces how intelligent the system is with what it can read, what it can answer, what it can execute, and who is responsible when it makes a mistake.

If the team still uses these terms loosely, align the vocabulary before any purchase or contract. The superintelligence glossary for executives helps separate concept, product and marketing promise.

What AI can do in citizen services

At a public service counter, a large share of questions are repetitive and do not require judgment: opening hours, the list of documents published in an official text, a deadline stated in a regulation, the status of a request checked in an existing system, and the correct channel for each type of demand. These are the cases where an AI agent is typically useful, provided it answers only from up-to-date official sources and states where the information came from.

The most important technical limit is the source limit. The agent should not fill gaps with general knowledge. If the answer is not in the material approved by the agency, the correct response is to say it does not know and hand the case to a staff member. This sounds obvious, but it is exactly the failure that shows up when a system is configured without a written scope.

What AI should not decide

The practical rule is to separate informing from deciding. Explaining a requirement is different from saying whether a citizen meets it. Describing a procedure is different from applying an exception. The list below is a starting point for the responsible committee to define the scope of each service:

  • Granting or denying a benefit, exemption, authorization or license must have a recorded human decision.
  • Interpreting exceptions, cases outside the rule and requests for review stay with a staff member.
  • Issuing a fine, sanction or collection notice must never be automatic.
  • Any mention of a health emergency, risk to personal safety or violence must be routed immediately to a human and to the emergency channel the agency has designated.
  • Sensitive requests, such as reports and requests involving personal data of third parties, should be received and forwarded by the agent without any assessment of their merits.
  • Denials of access to information need a written justification, checked by someone with the authority to sign it.

When a decision significantly affects the citizen and is made or supported by an automated system, the Brazilian data protection law (LGPD) provides a right to review. The topic is covered in the article on automated decisions under the LGPD, which should be read with the agency's legal team before the flow is designed.

Records and transparency: what to keep

Transparency without records is an empty promise. If the agency cannot reconstruct what the system told a citizen, it cannot correct errors, answer a complaint or prove that human review took place. Records must therefore exist from the first day of the pilot, and they serve both daily operations and oversight.

  • Channel, date and time of each interaction.
  • The text of the question and the text of the answer, marked as automated.
  • The official source used, with the version or date of the document consulted.
  • The reason for handing the case to a person, when applicable, in one objective sentence.
  • The name or department of the staff member who took over the case and when they did so.
  • The outcome recorded by the staff member, including whether the AI answer was corrected.
  • The access policy and retention period for records, defined by the agency under its internal rules and applicable law.

The notice to the citizen should appear at the start of the conversation and stay simple: it states that this is an automated system, explains what it can do, and shows how to reach a person. Teams working on WhatsApp should also follow the platform rules, as explained in the article on the 24-hour WhatsApp API window, because those rules limit what the agent can send after a citizen has been silent for a period.

Illustrative scenario: request triage at a city hall

Illustrative scenario, not a real agency. A municipal department wants to reduce the queue of repeated questions about documents and the progress of requests. The pilot covers only two services, using official texts published on the city website. The agent answers questions about required documents and checks request status in an existing system, with read-only access. When a citizen asks for an exception, reports an urgent situation or disputes a decision, the agent creates a summary of the case, sends it to the team, and tells the citizen the response time the department has set.

At the end of an agreed period, the team reviews a sample of conversations, checks whether the answers cited the correct source, verifies the handoffs, and decides whether to expand, adjust or stop. The decision rests on what the records show, not on a general impression. A similar setup for triage and deadlines is discussed in the article on superintelligence for notary offices.

How to set up a pilot with limits

The method we use follows four verbs: map, limit, pilot and hand over. It is not a special formula for the public sector, but it tends to work because it forces the team to ask the right questions before configuring anything.

  • Map the current flow, measuring the volume of each question type and who answers it today.
  • Limit the scope by writing down what the AI may answer, what it may only answer with a source, what it routes, and what it never does.
  • Name the owner of the official content and the owner of the operation, with names and roles.
  • Pilot one service, with complete records and a success metric agreed before launch, such as time to correct handoff or share of reviewed answers containing errors.
  • Hand over to the team with an operating manual, periodic review, and a clear rule for expanding, adjusting or stopping.

In practice, the pilot needs a test environment without access to real citizen data, and the agent's answers should be scored against a known set of questions before it goes into production.

Governance and references for managers

You do not need to start from scratch. The NIST AI Risk Management Framework organizes AI risk work around governing, mapping, measuring and managing risk across the system's lifecycle. The OECD AI Principles treat transparency, accountability and human oversight as central. In Brazil, the National Data Protection Authority (ANPD) publishes guidance on processing personal data, which should shape any flow that handles citizen information. These documents do not replace legal analysis of a specific case, but they give the manager a shared vocabulary and a checklist that control bodies recognize.

It is also worth remembering that an service agent is part of the public service, not a replacement for it. Agency service channels already follow rules on accessibility, language and deadlines, and the agent must respect them. The article on AI customer support offers a useful discussion of what to automate and what to keep with people, which applies well to public services.

Frequently asked questions

Can superintelligence be used today to serve citizens? Not as an entity that decides on its own, and no agency should plan around that idea. What exists are narrow AI systems configured for guidance, triage and routing tasks. Responsible use starts by defining what the system may answer and ends with human review of anything that affects rights.

Can an AI agent deny a request automatically? It should not. Denials, rejections and exceptions involve interpreting regulations and administrative responsibility, which requires a decision by a competent staff member. The agent can explain the procedure, state the deadline and record the request for human analysis.

How do you ensure transparency without making the conversation bureaucratic? The notice should be short and appear at the beginning: it states that this is an automated system, what it does, and how to reach a person. Behind the conversation, a full record of questions, answers, sources and handoffs makes it possible to account for decisions without asking the citizen for any extra steps.

Where should an agency start a project like this? Start with a service that has high volume of repeated questions and well-defined official sources, rather than trying to cover all service channels at once. Before any configuration, write the limits, name the owners and agree on the evaluation metric. Then run a pilot with complete records and decide, based on the data, whether to expand, adjust or stop.

Sources