Generative UI in public management: the manager asks, the panel is assembled

There is a scene that is repeated in almost every public body and in almost every data company. A manager needs an answer. He asks for a report. The request enters a queue, someone from the data area translates the demand into a query, sets up the panel, formats it, reviews it. Days later, the answer arrives, and it is often not exactly what the manager wanted, because the question changed along the way.

Generative UI attacks exactly this scene. Instead of asking and waiting, the manager asks in natural language and the interface is assembled instantly. "Show delayed processes by secretariat" stops being a call to the data team and becomes a dashboard that appears in seconds: bars by secretariat, table sortable by deadline, period filter.

It is a seductive promise for public management. It is also a terrain full of traps. I will treat both sides with the same seriousness.

The real gain: autonomy and decision speed

The first gain is autonomy. Today, managers depend on intermediaries to view their own data. Each new question reopens the queue. With Generative UI, part of this dependency disappears. The manager explores, filters, compares, without needing anyone to create the next vision.

The second gain is speed. Public decision has a window. A panel of health indicators that appears at the meeting, and not three days after it, changes the quality of the decision made there. Compressing the time between question and vision is compressing the time between doubt and action.

The third gain is less obvious: better use of technical teams. When the manager can answer routine questions alone, the data area stops becoming a reporting desk and goes back to doing the difficult work, modeling, data quality, in-depth analysis. AI does not replace the team, it removes what should not be in the queue from the queue.

In a digital government scenario, where the demand for transparency and rapid response only grows, this is powerful. But power without brakes is precisely the problem.

The risk that keeps me up at night: the wrong and convincing vision

Here is the most serious risk, and it is not technical, it is one of judgment. An AI-generated interface can be wrong and feel perfectly right.

A wrong paragraph raises suspicion, sounds strange. A wrong dashboard, with well-designed bars and aligned numbers, conveys authority. Visual form lends credibility to content, even when the content doesn't deserve it. This is the central danger of Generative UI in the public sector.

The model may misinterpret the question. "Late" according to what criteria, what base date, what definition of deadline? You can choose the wrong cut, aggregate improperly, cross sources that shouldn't be crossed. And it will deliver all of this on a beautiful screen, which the manager can take to make a budget or personnel decision.

The rule I defend is harsh: no generated interface should be treated as a source. It is a presentation layer on top of a source that needs to be reliable, auditable and versioned. If the data behind it is not complete, the beauty of the dashboard only makes the error more dangerous. Every important view needs to upload where it came from, with what filter, on what date.

Access governance: who can see what

In the public sector, data is not just numbers, it is information about people and sensitive decisions. Generative UI introduces a specific risk: if AI can assemble any view from any data, it may assemble a view that that user had no right to see.

The permission cannot live in the interface. It has to live in the data layer, before the model. The manager of a secretariat asks about delayed processes, and the AI should only be able to compose the panel with the data that that profile is authorized to access. Access control precedes screen generation, not after it.

This means that Generative UI in government only works on top of a data layer with serious permissions, by role, by agency, by sensitivity level. Without this, you don't have a management tool, you have a leak waiting to happen. The question “who sees what” has to be answered before the AI puts together the first screen.

LGPD and sensitive data: the legal limit

The governance layer meets the law. LGPD does not disappear because the interface became intelligent. On the contrary, it becomes more demanding.

Personal data processed by public authorities has rules of purpose and necessity. An AI that assembles views freely may, unintentionally, cross-reference information in ways that go beyond the original purpose of the collection. It may expose personal data in an aggregate that seemed harmless. In practice, it can create data processing that no one has evaluated.

Therefore, purpose and minimization need to be part of the path. AI cannot have unrestricted access to the collection just because it is convenient. The data catalog it can compose must be designed with the same legal discipline that applies to any public system that processes personal information, with a clear legal basis and a trace of who accessed what.

The audit record is also valid. Who asked, what did the AI assemble, with what data, at what time. In a public environment, the ability to be accountable for the use of information is not optional. Generative UI needs to be born auditable, or it shouldn't be born at all.

Continuity and dependence: thinking about the long term

There is a more silent, management risk. When the manager gets used to asking for everything in natural language, the organization starts to depend on an AI layer to see its own data. What happens when the model changes, becomes more expensive or interprets differently than before?

The healthy answer is not to throw away traditional paths. Official reports, institutional indicators, critical views continue to exist in a fixed and stable form. Generative UI comes in as a layer of exploration, agility, and long-tail questions. It speeds up everyday life, but it is not the only path to information that supports formal decisions.

Continuity, in public service, is a non-negotiable value. Governments change, teams change, suppliers change. The technology adopted needs to survive these changes, and that means not tying the house's intelligence to a single tool.

How to start without hurting yourself

If I had to recommend a path forward for a data agency or company, it would be gradual.

Start with non-sensitive data and low-risk questions, exploration and visualization, nothing that triggers automatic action. Treat each vision generated as a draft to be validated against the official source, never as the final truth. Build the permissions layer and audit trail first, because it's cheaper to do it earlier than to fix it later.

And keep people in the loop. Generative UI is an assistant to the data manager and the data team, not a substitute for the judgment of either of them. The screen assembles itself, but the responsibility for the decision remains human.

The promise is real: more autonomous management, faster decisions, technical teams freed up for what matters. The price is also real: reliable source, access governance, LGPD taken seriously and continuity thought out from the beginning. Anyone who faces both sides with the same honesty will get the best out of this technology. If you're evaluating this in your organization, start with the data layer and permissions, because that's where this story makes or breaks.