Who is IARA, CARF’s new member?

IARA has finally arrived at CARF. This is not a new member of the Council, but rather the long-awaited generative artificial intelligence system established to support adjudicatory activities, which was announced last year.

The Administrative Council of Tax Appeals (“CARF”) formalized the start of its activities through the publication of two ordinances: CARF Ordinance No. 142/2026, which establishes guidelines for the development and use of generative AI solutions, and CARF Ordinance No. 854/2026which establishes and approves version 1.0 of IARA, Artificial Intelligence in Administrative Appeals. In an official statement, the Council presented the measure as an institutional milestone in modernization, aimed at the responsible, safe, and ethical use of technology within the scope of administrative tax litigation.

According to the agency, IARA was launched to assist council members in drafting decisions, particularly in the search for case law references appropriate to the cases under review. The tool is entering its final phase of real-world testing with restricted access to a pilot group of 24 council members for a period of 30 days, developed by Serpro (Federal Data Processing Service) and curated by FGV.

The initiative is, in principle, positive, since CARF has not only formalized the use of artificial intelligence in its workflow but has also sought to surround this implementation with guidelines for human oversight, data protection, information security, and institutional accountability. In the official statement, the agency emphasizes that the tool aims to increase the efficiency, consistency, and technical quality of the adjudication process, without removing responsibility from the competent authority.

This regulatory framework is significant. The guidelines released by CARF highlight, among other points, a focus on the human person, the protection of personal data and confidential information, a reasonable duration of the process, and the need for effective, periodic, and adequate human supervision throughout the lifecycle of AI solutions. In other words, the technology was presented as a support tool, not as a substitute for the adjudicator’s decision-making activity.

Despite this, the development should be viewed with caution. In tax matters, especially regarding more complex issues, AI can identify semantically similar precedents without necessarily capturing, with the same precision, the specific legal reasoning of the dispute, the procedural context, or the precedent’s actual applicability to the specific case. Within the CARF context, this caution is even more relevant, because many disputes have a strong factual component and depend on a thorough analysis of the evidence presented and the particularities of each case; the discussion is not limited solely to legal rights. This risk is even consistent with the model adopted by CARF itself, which requires human review of the results generated by the tool.

In practice, this means that IARA can be quite useful for streamlining research and organizing case law references, but it does not eliminate the need for critical analysis. If the tool suggests grounds or decisions that do not exactly fit the subject under review, and this mismatch is reflected in the reasoning of the judgment, there may be increased scope for subsequent challenges and a backfiring effect, prolonging the discussion and leading t s through indirect channels, such as appeals for review of decisions, including motions for clarification and interlocutory appeals, which may represent a setback or even the risk of rendering fragile decisions if not granted.

This point deserves special attention, given that CARF’s Internal Rules of Procedure provide for the filing of motions for clarification when the judgment contains obscurity, omission, or contradiction between the decision and its grounds, or when there is an omission regarding a point that should have been addressed by the Panel, within five days of notification of the judgment. Furthermore, CARF’s own regulatory framework provides for motions for clarification with infringing effects, which reinforces the notion that significant flaws in the reasoning may, in theory, open the door to attempts to modify the judgment.

The same applies to the appeal (Agravo), which is applicable when the order admitting the Special Appeal denies or partially grants the appeal, and is often necessary when, upon a more superficial analysis, the minutiae of the specific case are not perceived in the precedent judgments—subtleties that perhaps IARA, despite having a proper name, is not yet equipped to address.

With these caveats in mind, the establishment of IARA is positive and represents a welcome advance, especially as it is accompanied by a formal concern for governance, security, and human oversight. In complex matters, AI can be a useful support tool, though it does not eliminate the need for careful reading, examination of the context, and rigorous scrutiny of the reasoning.