Sakana Marlin for Law Firms

Sakana Marlin is an autonomous research agent built by Sakana AI, a company based in Tokyo. It is not a chatbot and not a language model you hold a conversation with. You give it a research topic in plain language, and it runs on its own — for up to roughly eight hours per task — before returning a finished deliverable.

That deliverable is a comprehensive strategic report: an executive summary, strategic options, citations grounded in primary sources, and a set of presentation slides. Under the hood, Marlin uses AB-MCTS (Adaptive Branching Monte Carlo Tree Search), the method described in a NeurIPS 2025 Spotlight paper, to orchestrate several AI models and systematically explore and prune research paths. Sakana AI positions it for market analysis, competitive intelligence, M&A due diligence, and policy research.

This page explains where that shape of tool genuinely fits inside a law firm — deep legal and regulatory research, competitive and market research for the business of law, and background research for M&A legal due diligence — and, just as important, where it does not fit. It is not a drafting tool, not a redlining tool, and not a quick-answer assistant. We say so plainly.

What Marlin actually is

Marlin is a deep-research deliverable tool. The interaction model is simple: you describe a research question, and Marlin works autonomously for a long stretch — up to about eight hours — then hands back a finished report with an executive summary, strategic options, primary-source citations, and slides.

Because the runtime is measured in hours, not seconds, Marlin is designed for large research questions where you want a thorough, structured synthesis. It is not built for real-time back-and-forth. If you need an answer in the next minute, Marlin is the wrong tool by design.

• Input: a research topic in natural language.
• Process: autonomous run of up to ~8 hours using AB-MCTS to explore and prune research paths across multiple AI models.
• Output: a strategic report — executive summary, strategic options, primary-source citations, and presentation slides.

Where it fits a law firm

Three use cases map cleanly onto what Marlin does. First, deep legal and regulatory research: broad landscape questions where you want a structured survey rather than a single citation. Second, competitive and market research for the business of law — the kind of market analysis and competitive intelligence Sakana AI names as core purposes. Third, background research for M&A legal due diligence, which is one of Marlin's stated purpose-built areas.

In each case the value is the same: you hand off a big, well-defined research question and get back an organized deliverable hours later, rather than assembling it by hand.

• Deep legal and regulatory landscape research.
• Competitive and market research for firm strategy and business development.
• Background research for M&A legal due diligence.

Where it does not fit

Marlin is not a drafting assistant. It does not write your pleadings, briefs, or client emails. It is not a clause-by-clause contract review or redlining tool. And it is not a quick-question assistant — the multi-hour runtime rules out day-to-day Q&A.

For drafting, redlining, and fast interactive questions, pair Marlin with a different tool class — a chat or drafting model for quick work, or a dedicated contract-review tool for clause analysis. Marlin's job is the deep research that feeds those tasks, not the tasks themselves.

Confidentiality and privilege

Submitting a research topic to any third-party AI service raises duties under ABA Model Rule 1.6, which governs the confidentiality of client information. Before you put matter-specific facts, client names, or privileged material into Marlin, confirm what the vendor does with your inputs and outputs, review the data-handling terms, and consider whether the research question can be framed generically.

Many research questions — regulatory landscapes, market surveys, general precedent research — can be posed without disclosing confidential client detail. Prefer that framing. Where confidential facts are genuinely required, treat the decision as a confidentiality-and-informed-consent question, not a convenience question.

Limitations and mandatory human review

Marlin grounds its citations in primary sources, but it can still produce confident output that is wrong, including citations that do not support the proposition or that do not exist. This is not hypothetical: courts have sanctioned lawyers for filing briefs containing AI-hallucinated cases, most prominently in the Mata v. Avianca matter. A lawyer must independently verify every citation and every material assertion before relying on it.

Marlin is not a lawyer and gives no legal advice. Its output is a research draft to be checked, not an authority. Under ABA Model Rules 5.1 and 5.3, supervising lawyers remain responsible for work produced with the assistance of AI tools, including verifying that the underlying authorities are real, current, and on point — a step best done against Westlaw, Lexis, or the primary source itself.