Legal AI: plausible answer or source-grounded decision?

Anyone who has put a legal question to a generative AI tool knows the feeling. The answer arrives fast, well organized, with technical vocabulary and a confident tone. It reads like the work of a good associate. Sometimes it is. Other times the citation does not exist, the assumption is wrong or the conclusion ignores the exception that would decide the case.

The core issue is simple. Language models are excellent at producing plausible text. Legal work demands something else: a conclusion that can be checked, defended and signed by a professional who is accountable for it.

Plausible is not the same as grounded

A plausible answer sounds right. A grounded answer shows why it is right, where each statement comes from and under what conditions it would stop being true.

In daily practice, the difference shows up in concrete ways:

• The opinion cites a statutory provision but does not say which version of the text was considered.
• The analysis concludes that a tax credit is available but does not say this depends on a contract the client has not yet sent.
• The contract summary leaves out the amendment that changed the payment term.

None of these errors shows up on a quick read. All of them show up when the client, the tax authority or the other side reads the document carefully.

Four questions to ask any AI answer

Before using an AI-generated answer in legal work, put it through four questions. They apply to any tool.

1. Where did it come from? Every relevant statement should point to its source: the document and page, the statute, the decision. If it does not, you cannot check it without redoing the research.
2. What was assumed? Every analysis rests on assumptions about facts and classification. A good answer states them, so you can confirm or correct them.
3. What is unknown? Points without sufficient evidence should appear as uncertainties, not as conclusions in disguise.
4. Who reviews and signs? The answer must reach a qualified professional in a form that allows fast review, rather than going straight to the client.

If a tool does not help you answer those four questions, it may be useful for drafts, but not for decisions.

What Brazilian institutions have signaled

This is not only a technical debate. In November 2024, the Federal Council of the Brazilian Bar Association (OAB) approved recommendations on the use of generative AI in legal practice, built around four themes: applicable law, confidentiality and privacy, ethical legal practice and communication about the use of generative AI. The message for lawyers is clear: responsibility for the work stays with the professional, and the confidentiality of client information must be preserved.

In the judiciary, Brazil’s National Council of Justice (CNJ) issued Resolution No. 615 of 11 March 2025, setting guidelines for the development, use and governance of AI solutions in the courts. The text addresses human oversight, transparency and the classification of solutions by level of risk.

For lawyers and in-house teams working under Brazilian law, the direction converges: transparency about use, human oversight proportional to risk and accountability for whoever signs. On top of that come the obligations of the LGPD, Brazil’s data protection law, whenever personal data of clients or third parties goes into a tool.

How that translates into a product

Legal AI that takes these requirements seriously has to be designed differently from a general-purpose assistant. In practice, that means:

• Specialization by practice area. Tax law, contract review and portfolio management call for different analysis protocols. One assistant for everything tends to answer everything in the same tone.
• Evidence kept apart from inference. What came from a document or statute should appear separately from the model’s reasoning.
• Conditional conclusions. When the answer depends on an unproven fact, the conclusion should say so explicitly.
• A review point in the workflow. The analysis must reach the responsible professional in a form that allows checking, editing and approval, and only then move on.

That is how legal-tax intelligence in Castilho Torres IA works, with an assumption map, and how case law research works, with checking on the official court portal as an explicit step. The governance principles are set out on the technology and governance page.

A simple test for your team

Evaluating legal AI does not require a long project. A well-designed one-week test tells you a lot:

1. Pick three real questions your team has already answered, with supporting documents and no unnecessary sensitive data.
2. Put the same questions to the tool, with and without the matter context.
3. For each answer, apply the four questions: source, assumptions, uncertainty and review.
4. Measure how long the responsible professional takes to review the answer to a standard they would sign.

That last number is the one that matters. A tool that produces text quickly but requires long review does not save time. It just moves the time to the most expensive person on the team.

What changes when the answer is grounded

When AI delivers assumptions, sources and uncertainty in an organized way, review stops being a full reread and becomes a targeted check. The partner spends time on what is genuinely contested. The associate learns from visible reasoning. And the client gets an answer that says clearly what is proven and what still depends on a document.

That is the standard we propose for any legal AI: it is not enough to sound right. It has to be checkable.