What happened
Brazil's Administrative Council of Tax Appeals (Carf) decides appeals on federal taxes. In January 2019 it held about 122,000 appeals worth more than R$600 billion. From 2019 Carf used artificial intelligence to group appeals that raise the same legal thesis into batches. One appeal per batch was drawn as the paradigm, and its decision applied to the rest. The Federal Court of Accounts (TCU) audited Carf from November 2019 to April 2020 and ruled on 24 February 2021. The auditors found that in 2019, for the first time in the period studied, Carf's panels decided more appeals than arrived. As one possible explanation they named a fall of nearly 20% in new appeals, together with the AI batching. The panels issued 22,404 decisions in 2019, up 107.50% from 10,797 in 2017, while reported hours rose 35.73%. Yet the stock fell only 1.06%, which the auditors linked to a near doubling of special appeals to the Superior Chamber. The TCU did not measure the AI's part. It neither names the tool nor shows that it learns from data.
Sources
-
Audit report, by authors with no tie to the system.
somente em 2019, as TO/TE julgaram mais processos do que as entradas. Uma possível explicação para esse fenômeno seria a redução em quase 20% na quantidade de processos novos vindos das DRJ em relação ao anterior (v. tabela 24 do apêndice G), bem como a utilização de inteligência artificial para formação de lotes de processos para julgamentos repetitivos
Details
AI system
- Workflow
-
From 2019 Carf used artificial intelligence to find appeals on the same subject and form batches with the same legal thesis. One appeal in each batch was drawn as the paradigm, and its decision was applied to the rest of the batch. Carf did not suspend judgments to wait for a thesis to be settled, a step that the courts take.
Follow-up
- Conditions for replication
-
The appeals in a batch must share the same legal thesis. Decisions of the panels must become final for the stock to fall, since special appeals to the Superior Chamber keep decided cases in it.
Entry details
- Sector
-
Justice administration: administrative tax appeals.
- Purpose
-
A developer that the audit does not name built Carf's AI tool to find appeals on the same subject and group them into batches for repetitive judgment. Carf adopted it from 2019 to reduce its stock of appeals.
31 more questions not yet checked
model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, AI contribution, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.
Cite this record
Peter H. Vartanian. (2026). Audit court cites AI batching as a possible reason for Carf's output (AIBD-0995). Artificial Intelligence Benefits Database, edition of 1 November 2026. https://petervartanian.xyz/aibd/records/AIBD-0995.html (archived at https://doi.org/10.7910/DVN/AADSII, version 7.0)