In under a decade, Brazil solved a problem many wealthy countries still have not: it brought federal public services into one place, under one digital identity. The question now is a different one.

There is a cycle of digital government that can reasonably be considered closed. It began with digitising forms, moved through integrating databases and ended in the single platform, with a digital identity of national reach and defined trust levels. It was an institutional effort of rare scale, and it worked.

The mistake now is to treat the next cycle as a continuation of the last. It is not. Digitising more services under the same logic produces diminishing returns. What shifts the level is something else.

Three shifts that define the next cycle

From the requested service to the anticipated one

In the current model, citizens must know they have a right, know the name of the service and know where to request it. Three barriers before the first interaction. The next level inverts the initiative: the administration recognises eligibility and offers. That requires data integration with a clear legal basis, not merely a better interface.

From the interface to the conversation

A public service designed as a form presumes a citizen able to translate their need into the structure of the state. Most cannot, and should not have to. The conversational, multimodal layer changes the nature of the interaction: the citizen describes the problem in their own words, and the state does the translating. This is technically viable today. What is missing is institutional decision and governance.

From task automation to agent orchestration

Automating a task reduces cost. Orchestrating agents that carry out whole chains of work changes the design of the operation, and with it the design of control. This is where the technical conversation has to become a governance conversation, because an agent that acts also errs, and error in a public service carries legal and social consequence.

What the next cycle lacks is not technology. It is knowledge management, data governance and named accountability for what the machine starts deciding.

The real bottleneck is not the model

Anyone trying to put an agent into production inside a public organisation discovers quickly that the language model is the easy part. The agent fails because the correct sequence was never written down: it lives in the heads of three experienced people, and each does it differently. This is not an absence of information. It is an absence of codified knowledge.

Building the corporate knowledge base that sustains agents is a management project, not an infrastructure one. Every data source to be integrated is a small project with an owner, a rule and an update cycle. In an organisation with dozens of silos, that does not scale through central effort: it scales when each area learns to build its own layer and connect it to the whole.

What senior leadership must decide now

  • Whether a corporate model is under contract, or whether AI use will carry on happening outside the institutional perimeter.
  • What the usage policy is: what is permitted with public data, what is prohibited with sensitive data, and who answers for the difference.
  • Who is the named owner of every agent in production, and against which indicator they are held.
  • How the organisation measures adoption today, so it can know whether next year's investment improved anything.

None of those decisions is technical. All of them need to be taken before the next procurement, not after. The institution that decides first does not win by being more modern. It wins by not having to undo.