Brazil: how the Chamber of Deputies uses AI across high-volume legislative tasks

The problem

The Chamber of Deputies is the lower house of Brazil's National Congress, the body that drafts and passes the country's federal laws and publishes legislative information to the public. Its work runs at a scale that is hard to manage by hand. The office that helps draft legislation takes in tens of thousands of requests a year; amendments arrive in large numbers during sessions; hours of spoken debate have to be transcribed for the official record; and the Chamber's public website is used on a very large scale.

Several of these tasks carried friction. Directing each drafting request to the right group of specialists depended on individual judgment and could be uneven. Checking whether a proposed law resembled existing studies or bills was painstaking. Transcribing debates by hand was slow. And the number of public comments submitted on bills was large enough that drawing clear conclusions from them was difficult.

The solution

To take on this range of work, the Chamber built an integrated set of AI tools it calls Ulysses, named after Ulysses Guimarães, the parliamentary leader who presided over the assembly that wrote Brazil's 1988 Constitution. The system was designed as a suite of modules, each handling a part of the legislative process and aligned with the Chamber's wider digital strategy.

The modules are divided across the workflow. On the public side, one module organises and reorganises legislative content on the Chamber's portal by theme, so citizens can find material on a subject in one place. Inside the drafting process, one module routes incoming requests to the most suitable group of legislative counsel, the lawyers who help draft legislation, and offers ranked suggestions that specialists check before they go out; another reads each request and finds existing studies or proposals that are similar in meaning, which helps avoid duplication; and a third analyses amendments during a Bill's passage, working out where each belongs and grouping ones that overlap. A transcription module turns spoken proceedings into text for the record. A public-participation module reads the written comments citizens leave on bills through the Chamber's online polls and groups them by argument, so the range of opinion can be taken in at once. Two further modules use facial recognition: one to authenticate members taking part in remote votes, and one, limited to official photographs and noted as under review, to identify members in images from public meetings.

Human oversight was built into the design. The request-routing module is semi-automated, with specialists reviewing its suggestions before they go out, and the transcription module pairs the automated draft with human editing before anything becomes an official record.

The programme grew over more than a decade, from an early accessibility tool around 2013 to the 2018 launch, and in December 2025, the Chamber turned it into an institution-wide AI programme with coordinated projects across the House. That phase added AI and data governance policies that set out who is responsible for what and how the technical teams coordinate, an internal assistant, Ulysses Chat, that answers staff questions about the House's own rules and procedures, staff access to external models such as Claude, Gemini and GPT, and tools that help sort proposals, send matters to the right committees, and point up possible clashes with the constitution, all with people in review.

The impact

1. The suite spans the entire legislative process.
   Its modules touch the public website, the drafting and amendment process, the transcription of debates, the analysis of public comments, and the authentication of voting, which together cover much of the legislative process.

2. It operates at a large scale.
   The legislative counsel office handles more than 20,000 requests a year for the request-analysis module to process, and at launch, the Chamber's website was already drawing around 200 million visits a year. These figures show the volumes the system handles. No efficiency or accuracy gains have been measured or published.

3. The Chamber now treats Ulysses as a long-term institutional capability.
   Presenting the 2025 phase, its director-general voiced an ambition for the House to become a worldwide reference point among parliaments for its use of AI, and its secretary-general stressed the need to maintain human review at all stages.

Lessons learned

• The distinctive choice was an integrated suite tied to a digital strategy. The modules were built to fit the Chamber's own rules and traditions, for example, keeping a draft confidential until it is formally put forward, and integrating transcription with the Chamber's existing stenography, the shorthand record of proceedings. A government attempting this would need a sustained digital strategy and in-house technical capability, more than an off-the-shelf purchase.
• The design kept a person responsible for each output. Across the modules, the same principle holds: the system proposes, and a person decides. The leaders presenting the 2025 phase placed human review at the centre, and the programme's AI lead has described human oversight and refinement as central to its success.
• Governance was made explicit and sits apart from the national AI law. The programme rests on data governance and ethical AI policies, further formalised in the December 2025 governance documents. These are the Chamber's own internal rules, separate from Brazil's proposed national AI law, which is still moving through Congress.

This case study was written with assistance from artificial intelligence.