How AI Is Helping Government Companies in Brazil Cut Costs and Improve Efficiency

AI is already helping Brazilian government companies squeeze costs and speed service: the National Treasury's text-classification model slashed subnational expenditure tagging from 1,000 hours to 8 hours with over 97% accuracy - a dramatic efficiency gain detailed in the IMF's report on STN's work (IMF report on STN AI fiscal transparency in Brazil), while World Economic Forum pilots show how an “AI Procurement in a Box” approach enabled São Paulo Metrô to scope predictive‑maintenance buys and Hospital das Clínicas to build a data roadmap for safe AI adoption (World Economic Forum case study: AI Procurement in a Box in Brazil).

These wins sit beside tougher realities - 11,000 government transparency portals and pilots like a WhatsApp-based TransparencIA that scored 75% accuracy - so practical training matters: Nucamp's 15‑week AI Essentials for Work course teaches prompts, tools and workplace use cases to help public servants apply AI responsibly (Nucamp AI Essentials for Work syllabus).

“The goal isn't to take the place of human beings but to get the court's work done faster so as to increase human productivity. The really exhausting and repetitive work can be done by computers,” Lotufo said.

Brazil's public sector is no longer just experimenting - ministries and state governments are scaling AI into core services, backed by national strategies and big money: the PBIA and EBIA steer policy and capacity building while investments in AI and generative projects are forecast to exceed BRL13 billion (about USD2.4 billion) by 2025 and the PBIA earmarks roughly BRL23 billion through 2028, signaling serious national commitment (Brazil Artificial Intelligence 2025 trends and developments - Chambers Practice Guide).

Practical deployments span tax analytics, health imaging, predictive maintenance for transport, chatbots for citizen services and document automation - about 37% of public agencies reported AI use in 2023 - even as regulators like the ANPD tighten rules on data and platforms to protect privacy and fairness (Brazil AI market growth, adoption statistics, and regulatory landscape).

The result is a mixed but fast-moving picture: measurable efficiency gains and tighter governance, with public debate - for example over facial recognition and surveillance - ensuring that speed doesn't outpace safeguards.

The INSS “Isaac” rollout is a vivid lesson in why efficiency and equity must travel together: Dataprev's automation cut queue times dramatically - death pensions could be recognised in as little as 12 hours and automated decisions rose from ~17% in 2022 to over 36% in 2023 - but those gains came with a sharp rise in automatic rejections and new barriers for the most vulnerable.

Audits show more than 800,000 automatic denials in 2022 (a 300% jump versus 2021), appeals jumped into courts and boards, and CNIS data quality problems (over 24 million inconsistent entries flagged by TCU) meant the system sometimes punished paperwork errors rather than fixing them.

Workers and unions warned that a person's retirement was once rejected in six minutes, while governance gaps - only a handful of staff devoted to automation and no clear public justification or co‑design - left citizens with little recourse.

The balanced critique in the Internet Policy Review analysis of INSS “Isaac” automation and the CGU audit report on INSS automation (EAUD) make the point plainly: automation can speed outcomes, but without transparency, safeguards and stronger data governance those time savings can translate into denied rights for people who can least afford them.

“The automated analysis of benefit requests is one of the actions that Social Security has adopted to reduce the response time for citizens requesting a service or benefit.”

In a standout systems win, Brazil's National Treasury (STN) turned a brutally slow, manual COFOG tagging task - roughly 100,000 subnational budget records a year - into a near-instant process by deploying a probabilistic text‑classification pipeline built on Convolutional and Recurrent Neural Networks; the result was jaw‑dropping: classification time fell from 1,000 hours to just 8 hours (a 12,400% efficiency gain) while accuracy topped 97%, enabling the 2024 “Despesa por Função do Governo Geral” publication and invitations to international forums (IMF analysis of STN COFOG classification and fiscal transparency in Brazil).

That technical choice - CNNs and RNNs for textual coding - echoes broader findings on neural architectures for text and report classification (comparative CNN and RNN study for medical text classification (PubMed)), and STN is already scaling the method to climate‑tagging with the IDB, illustrating how smart model selection plus rigorous workflows can turn months of work into hours while strengthening fiscal transparency across Brazil.

The São Paulo Metrô pilot shows how a pragmatic, in‑house AI approach can move predictive maintenance from promise to everyday operations: technicians built an Asset Monitoring System (AMS) that watches escalators, lifts, trains, tunnel ventilation and power supply systems across lines 1‑Blue, 2‑Green, 3‑Red and 15‑Silver, streams that data to the Metro's Maintenance Control Centre at Jabaquara, and raises alerts so operators can dispatch crews or give remote support before a fault hits service - literally preventing a passenger disruption rather than reacting to it (São Paulo Metro AI predictive maintenance system - RailJournal).

That operational model echoes other Brazil pilots - like thyssenkrupp's MAX elevator service, which cut elevator downtime in pilot sites by about half - illustrating how targeted sensors plus AI decision layers turn noisy signals into timely, cash‑saving interventions (MAX predictive elevator maintenance - TiInside),

“one alert, one crew”

a vivid detail that explains why metros and agencies view PdM as an immediate, operable efficiency lever.

Hospital das Clínicas in São Paulo turned a sprawling, mission‑heavy environment - teaching, research and clinical care fed by more than 60 systems and protocols - into a practical roadmap for safe AI adoption by prioritizing data integration, workforce training and harmonized architecture: the team built a data lake to bring internal and external registries together, invested in technical training and agreed data standards, and sketched next steps for datamarts and feature stores so models can be reused and audited (see the World Economic Forum case study: AI Procurement in a Box - Hospital das Clínicas roadmap World Economic Forum case study: AI Procurement in a Box - Hospital das Clínicas roadmap).

That approach mirrors best practice in health informatics - where data lakes, datamarts and feature stores play complementary roles in the reuse pipeline (JMIR review: data lakes, datamarts and feature stores in health informatics) - and echoes real Brazilian experience with hospital teams speeding research workflows using platforms like Dataiku in Porto Alegre, which highlights how tooling plus governance creates reproducible, auditable AI pipelines (Dataiku case study: Hospital de Clínicas de Porto Alegre streamlining clinical research workflows).

The vivid payoff: stitching 60+ siloed systems into one governed data backbone turns months of manual wrangling into consistent, reusable inputs for safe, cost‑saving AI pilots.

Buying AI in Brazil means more than issuing a specs sheet - it requires tailored procurement, governance and capacity building so public agencies actually capture value while managing risk; the World Economic Forum's AI Procurement in a Box toolkit, adapted by C4IR Brazil, shows how ten practical guidelines, stakeholder engagement and even Brazil's first-ever Algorithmic Impact Assessment helped São Paulo Metrô translate a predictive‑maintenance ambition into a procureable, mitigated project (the Metro used innovation procurement rules to address 1,840 minutes of service interference recorded between 2016–2020) (World Economic Forum: AI Procurement in a Box - Brazil case study).

Pilots also proved why hospitals and metros must invest in data maturity (Hospital das Clínicas' work to unify 60+ systems) and why governments need monitoring indicators and workforce training to oversee contracts responsibly (Implementation insights for AI Procurement in a Box - Glass Community).

Start with inventories, algorithmic impact checks and clear evaluation criteria to turn promising AI pilots into durable, auditable public services (Operational AI controls and inventories for government AI projects).

Measured gains from AI in Brazil's public sector are tangible but must be read alongside offsetting costs: INSS's Atestmed alone is forecast to cut sickness‑benefit spending by about R$2.58 billion in 2026, saving roughly R$980 per grant by replacing costly in‑person exams with document analysis (INSS expects to save R$2.5bn with Atestmed), but those headline numbers sit next to real liabilities: automated denials surged (hundreds of thousands of rejections and appeals), driving judicialisation and service frictions that erase some efficiency gains and, in extreme cases, produced stories of a retirement claim denied in six minutes (Analysis of INSS automation and its social costs).

Meanwhile, the discovery of more than R$6.3 billion in undue discounts points to additional fiscal and reputational offsets that can outweigh short‑term savings; the lesson is clear - documented efficiency must be paired with governance, monitoring and appeals capacity so time‑saved doesn't become rights‑denied.

“It's a humiliation,” de Brito said.

Risk is front‑and‑centre for Brazil's public‑sector AI: independent audits by the CGU and TCU expose weak governance, tiny automation teams and brittle data foundations that turn speed into a hazard rather than a public good.

Faulty CNIS records (over 24 million inconsistent entries) and a lack of traceability mean automated pipelines routinely fumbling real lives - automatic denials jumped sharply in 2022 (more than 800,000 rejections, a ~300% rise) even as automatic decisions climbed into the tens or hundreds of thousands, driving appeals, judicialisation and reputational damage; one high‑profile case reported a retirement claim denied in six minutes, a vivid sign that throughput isn't the same as fairness.

Audits also flag a missing public justification, scarce monitoring indicators, no clear risk‑acceptance threshold and little external validation, so decisions remain a black box for citizens and control bodies alike.

Policymakers and project leads must pair automation with stronger data curation, accountability metrics and citizen‑facing remedies, echoing detailed critiques in the Internet Policy Review and the official CGU audit that call for governance, transparency and participatory checks before efficiency becomes harm (CGU audit report on INSS automation, Internet Policy Review analysis of INSS “Isaac” automation).

“The automated analysis of benefit requests is one of the actions that Social Security has adopted to reduce the response time for citizens requesting a service or benefit.”

Beginners working on AI in Brazilian government should start with very small, defensible steps: map every AI system and run a preliminary risk classification so high‑impact tools are flagged early (this is a cornerstone of Brazil's emerging regulatory approach and practical compliance playbooks - see Nemko's guide to AI governance in Brazil AI Governance Brazil: Navigating Policies & Compliance); next, invest in data quality, lineage and a simple catalogue because

“data governance is the concrete foundation and AI governance is the roof”

- without that foundation models will amplify errors and bias (the four‑pillar approach to data + AI governance is a useful starter framework: data, lifecycle, operations and ethics AI and Data Governance: The Essential 4‑Pillar Framework).

Require iterative algorithmic impact assessments and public participation for high‑risk uses, keep human‑in‑the‑loop review paths, bake monitoring and model cards into contracts, and prefer sandboxed pilots with clear evaluation criteria so a successful time‑save doesn't become a rights‑denying mistake; civilsociety engagement and transparency - already stressed in Brazil's legislative debate - make deployments more robust and publicly defensible (Lessons on public voice in Brazil AI regulation).

These basics create a low‑cost, high‑trust starting point for scalable, auditable public‑sector AI.

The future of AI for Brazil's government companies looks both promising and conditional: the Brazilian Artificial Intelligence Plan (PBIA) commits roughly R$23 billion through 2028 - including a planned supercomputer ranked among the world's top five - to build sovereign infrastructure, boost R&D, and scale practical public‑sector uses from health to fiscal transparency (ABES - Brazilian Artificial Intelligence Plan (PBIA) full text); the PBIA's goals are echoed in official timelines and highlights that stress ethics, social inclusion, and concrete service gains (LNCC - Final version summary of the Brazilian Artificial Intelligence Plan (PBIA)).

Success, however, will hinge on execution - strong procurement, data quality, clear regulatory guardrails and a trained workforce so efficiency gains don't become rights‑denying errors - and that's where targeted, practical training matters: short, workplace‑focused courses like Nucamp's AI Essentials for Work teach prompt design, tool use and governance basics that help public teams move pilots into safe, auditable operations (Nucamp AI Essentials for Work syllabus (AI training for public-sector teams)).

The PBIA is a call to action: with infrastructure, governance and people aligned, Brazil can turn national scale data and the SUS's reach into safer, faster, and more equitable public services.

“It is essential that Brazil does not sacrifice democratic safeguards in the name of technological competitiveness.”