evidoria

← Back to browse

Good practice Imported

ChatTCU — Brazil's Federal Audit Court Logs 1.3 Million Internal AI Chats, With Mandatory Human Review

Brazil · Brasília · See the Brazil profile · See the Brasília profile

Evidence: Descriptive / self-reported Top 22% 73/100 · Ask Evidence Copilot about this practice

Brazil's federal audit court built ChatTCU, an internal generative-AI assistant on Azure OpenAI with retrieval over its own case law. By June 2026 it logged 1.3 million chats among 3,860 staff, with mandatory human review of every output before use.

1.3 million
Chats processed (Feb 2023 - Jun 2026)
4.9 million
Messages processed (Feb 2023 - Jun 2026)
3,860
Distinct users (as of Jun 2026)
10,000-30,000
Typical daily message volume (as of Jun 2026)
ChatTCU — Brazil's Federal Audit Court Logs 1.3 Million Internal AI Chats, With Mandatory Human Review

Details

Maturity
Established
Promoter
Tribunal de Contas da União (TCU)
Period
February 2023 – June 2026
Keywords
public audit, internal government operations, generative AI

Context

The Tribunal de Contas da União (TCU), Brazil's federal audit court, needed a way to give staff natural-language access to institutional knowledge (case law, prior rulings, administrative records) that general-purpose AI models cannot see.

Objectives

Give TCU staff a generative-AI assistant to search institutional knowledge, draft preparatory text for opinions and rulings, and support audit and financial analysis, while requiring human review of every output.

Activities

TCU launched ChatTCU on Azure OpenAI (GPT-3.5) in February-March 2023, released a second version in June 2023, and added retrieval-augmented generation (RAG) over TCU's own case law and administrative records in a third version in October 2023. Access is restricted to TCU personnel over an encrypted, non-public network link, and every AI-drafted output must be reviewed by a human before use.

Results

By June 2026, ChatTCU had processed 1.3 million chats and 4.9 million messages among 3,860 distinct users, with typical daily volumes of 10,000-30,000 messages. TCU has reportedly begun open-sourcing the tool's code for other public bodies (July 2026).

Conclusions

ChatTCU shows sustained, growing internal adoption inside a major audit institution and documented governance safeguards, but the published figures describe usage volume, not an independently measured effect on accuracy, error rates or staff time saved.

Implementation
Indicative cost
Medium (€50k–€500k) — Built on commercial cloud AI infrastructure (Azure OpenAI/GPT-3.5) with incremental version releases (2023) rather than bespoke model development; ongoing costs likely scale with usage volume and dedicated network/security infrastructure.
Time to results
Long (> 3 years) — Iterative rollout over three versions from February 2023 to October 2023, with continuous operation and growing usage reported through June 2026 (over three years).
Staffing & skills
TCU IT Secretariat (led by IT Secretary Rainério Rodrigues Leite and Deputy Technology Secretary Wagner Miranda Costa), TCU auditors and legal staff as end users, required to review every AI output

Conditions for success

  • Mandatory human review of every AI-drafted output before use
  • Retrieval-augmented generation over the institution's own case law/records rather than relying on general model knowledge
  • Restricted access, encryption at rest/in transit, and a dedicated non-public network link

Common failure modes

  • No independent evaluation of accuracy or error rates has been published, so quality risks in AI-drafted legal text are not externally verified
  • TCU's own domestic AI-maturity survey and OECD commentary note most federal Brazilian institutions remain at early AI-adoption stages, making TCU's advanced use an outlier that may be hard for less-resourced peer institutions to replicate

Where it fits

Governance type
national supreme audit institution
Scale
national, ~3,860 users
Income level
upper-middle-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful