Can generative AI make public finance truly accessible?

Brazil has achieved something extraordinary. The country maintains 11,000 government transparency portals – one for nearly every public entity – yet most citizens find them incomprehensible. It's like giving everyone a calculator when they can't do sums. This paradox, explored at the Brazil Forum UK on 14-15 June 2025, illuminates a challenge facing democracies worldwide: can artificial intelligence finally make public spending data accessible to citizens?

The statistics are striking. Brazil ranks 7th globally on the Open Budget Index, ahead of Germany and Canada. Yet our research reveals that barely one-third of users can make sense of the data they find.

The transparency challenge

Following major corruption scandals, Brazil mandated in 2009 that all 5,570 municipalities publish real-time spending data online. The scale is staggering: from São Paulo, with budgets rivalling small nations, to remote Amazonian villages – each must maintain sophisticated financial websites. According to ATRICON (National Association of Audit Courts), this has resulted in over 11,000 separate portals.

This challenge extends globally. India's Panchayat accounts, UK council spending data, and Nigerian state budgets face similar problems. Governments have embraced transparency as standard practice, yet citizens remain distant from public finances.

The economic case for transparency is well-established. World Bank analysis of 38 studies confirms that openness reduces corruption and borrowing costs whilst improving public services. South Korean experiments demonstrated budget transparency cutting waste by 23 per cent. Brazilian municipalities with participatory budgeting saw tax revenues rise by 16 per cent and infant mortality fall by 18 per cent.

However, transparency without comprehension achieves little. The gap between data availability and citizen understanding remains vast across democracies.

Testing an artificial intelligence solution

Could artificial intelligence bridge this gap? The TransparencIA pilot, developed through collaboration between SuperDash Software and Espírito Santo's transparency agency, explored this possibility. The system was designed for WhatsApp deployment – essential in Brazil where WhatsApp penetration far exceeds email usage.

Researchers conducted controlled tests to evaluate the proof-of-concept system, systematically querying GPT-3.5's ability to interpret citizen questions and retrieve accurate fiscal data. The methodology involved academic literature review followed by precision testing across various query types and complexity levels.

The economics have evolved considerably. The 2023 pilot projected costs of $2,000 monthly to serve 20,000 users. Current models offer superior performance at lower costs, whilst open-source alternatives raise possibilities for data sovereignty – crucial for government applications.

Confronting limitations

The results proved sobering. The pilot achieved 75 per cent accuracy – acceptable for consumer recommendations but problematic for public finances. Even if current models reach 95 per cent accuracy, questions remain: what error rate is acceptable for government data? Who bears responsibility when AI misreports spending?

More fundamentally, only 22 per cent of Brazilian state portals maintain consistent, reliable data. Sophisticated analysis cannot compensate for poor data quality – a challenge from Stockholm to São Paulo.

Institutional barriers compound technical challenges. Government staff express concerns about job displacement. Politicians accustomed to selective transparency resist comprehensive openness. Across Brazil's federal system, these obstacles multiply through thousands of independent entities, each protecting established practices.

The core tension remains political rather than technical. Transparency serves different constituencies differently: reformers seek accountability, politicians ensure compliance, bureaucrats minimise risk. Technology cannot resolve these fundamental conflicts.

Lessons for implementation

Brazil's measured approach offers valuable guidance. Rather than rushing deployment, officials recommend phased implementation: internal validation through rigorous accuracy testing, limited trials using innovative procurement frameworks, then scaled deployment with clear accountability structures.

Throughout, data protection and algorithmic transparency remain non-negotiable. Democratic applications demand higher standards than commercial ones – when AI misreports government spending, public trust erodes.

The international collaboration between Brazilian transparency experts and Indian technologists suggests possibilities for South-South cooperation. Yet success requires more than technical solutions. Quality data infrastructure, institutional capacity, and sustained political commitment prove equally essential.

Implications for democratic governance

As governments worldwide explore AI deployment, Brazil's experience offers timely insights. Data infrastructure must precede AI implementation. Federal complexity multiplies every challenge – solutions must accommodate vast disparities in capacity. Institutional change proceeds slowly. Technology procurement happens quickly; transforming bureaucratic culture requires sustained effort.

The potential remains compelling. Accessible public finance data could transform democratic participation, expose corruption, and improve service delivery. Yet achieving this requires governing AI to serve democratic purposes rather than efficiency alone.

For policymakers globally, Brazil's journey provides both caution and inspiration. Careful experimentation, rigorous evaluation, and patience prove more valuable than technological enthusiasm. In an era of declining public trust, making government genuinely comprehensible to citizens represents one of democracy's most pressing challenges.

The question isn't whether AI can decode government spending – it's whether democracies can accept that technological capability without political will achieves nothing. Brazil's experiment reveals an uncomfortable truth: the barriers to transparency aren't primarily technical or even educational. They're embedded in institutions that benefit from complexity, in federal systems that fragment accountability, and in the gap between compliance and genuine openness. Perhaps the real innovation isn't teaching machines to read budgets, but creating incentives for governments to want citizens to understand them. Until then, those 11,000 portals remain monuments to transparency's promise – and its limits.

This analysis draws on research presented at the Brazil Forum UK, June 2025. The author's policy brief is available on the BSG website.