AI can already do the legal work. The rules are still India’s to write

In April, a Magic Circle firm put an AI model in front of 5,700 of its people. Freshfields did not run a quiet pilot in one team. It gave the whole firm access to Claude, the model built by Anthropic, across 33 offices. In the first six weeks, the firm said, use of it went up by about 500 percent.

A month later, Thomson Reuters, the company behind Westlaw, said it had rebuilt the next generation of CoCounsel, its main legal assistant, on Anthropic’s agent toolkit, while keeping CoCounsel multi-model. Thomson Reuters says about a million professionals in 107 countries have chosen CoCounsel. It now reads across 1.9 billion Westlaw and Practical Law documents through Claude.

This was not a one-off experiment. In May, Anthropic released a legal product of its own, Claude for Legal, with tools built for a dozen practice areas and secure connections into the document and practice systems firms already use. More than 20,000 legal professionals signed up for a single Anthropic webinar on how to use it, the largest legal session the company had run. At Quinn Emanuel, a partner built the firm’s litigation platform on Claude. Holland & Knight put it to work across litigation matters. Harvey, the best known legal-AI company, says around 80 percent of the hundred largest US firms now use its product, and it runs Claude as one of its models.

I work in legal technology. For years, the honest answer to whether AI could really do legal work was simple: not yet, and be careful. That answer has changed. The debate about whether these tools work is mostly settled. The open question now is who decides how lawyers are allowed to use them. The United States is answering it by doing. India gets to answer it on its own terms, and the choice is live right now.

Let me start with why Claude, and why lawyers are the ones adopting it.

The first reason is that it shows its work. Anthropic built a feature it calls Citations that makes the model tie each answer back to the exact sentence in the document it read, so a lawyer can check the source before relying on it. That matters in a profession where an unchecked assertion is a professional risk, not a small one. A tool that points to where its answer came from is a tool you can supervise. A tool that just sounds right is not.

The second reason is the pace. In the last week of September 2026 alone, Anthropic shipped two new flagship models six days apart, Opus 5.5 and Sonnet 5.5. It had released Opus 5 in July. By May this year investors valued the company at about 965 billion dollars, up from 183 billion the previous September. I do not cite that as proof the technology is any good. Money is a bet, not a result. I cite it because that kind of capital buys a release schedule most law firms cannot picture, and the models keep getting better every few weeks.

Anthropic also writes down how it tries to keep the model in check. It trains Claude against a published set of rules it calls a constitution, and it runs a public safety policy that it has revised nine times, the last in July 2026. For a profession that answers for every word it files, a vendor that writes its guardrails down is easier to trust than one that keeps them in a drawer.

I use these tools most days, so let me be plain about what they get right and what they miss. Hand a current model a 200-page facility agreement and it will pull every change-of-control and termination clause faster than any junior I have worked with, and summarise them cleanly. Ask it a settled question of law and it will usually answer well. The failure is narrower and more dangerous than people expect. Every so often it will produce a citation that reads perfectly, with the right court and a plausible year, for a case that does not exist, and it states the invented one with exactly the same confidence as the real ones. That is the risk in a single line. The mistakes look just like the good work.

This is not a worry on paper. On the Legal Agent Benchmark, built by the legal-AI company Harvey and published as a live leaderboard by Vals AI, the best agents complete only about one full multi-step task in four without an error, on the leaderboard as it stood at the end of September 2026. In a US bankruptcy matter this year, the firm Sullivan and Cromwell filed documents containing citations the AI had invented, opposing counsel caught them, and a partner wrote to the judge to say the firm deeply regretted the errors. An authoritative wrong answer is worse than no answer, because it is the one a busy lawyer is least likely to catch. These tools are good enough to trust with a first draft. They are not safe to file without a human reading every line.

The United States is settling these questions the way it settles most things, by doing, live, inside its firms and courts, and cleaning up the mistakes. India does not have to follow that order. It can decide the rules while the tools are still arriving. And the striking thing is that our courts are not waiting for anyone. They are already deciding.

The Kerala High Court wrote the first binding court policy on AI in July 2025 and kept public chatbots out of any part of judicial decision-making in its district judiciary, on confidentiality grounds. A year on, the Supreme Court went further. In July it set aside two tribunal orders because they rested on case law an AI had made up, six citations, fabricated or unverifiable. The bench of Justices Narasimha and Aradhe declared that it is misconduct for an advocate to cite such judgments without verification, and directed the Bar Council of India to convene a committee, frame a guiding principle and set out the disciplinary action that follows a breach. No advocate had cited the fabricated authorities in that case, because the tribunal had found them through its own research, so the rule the Court stated is a declaration for the future. That is Pooja Ramesh Singh v. Jammu and Kashmir Bank, 2026 INSC 668, decided on 2 July 2026. Six months before it, the Bombay High Court had already put a price on the same mistake, 50,000 rupees in costs on a litigant who filed submissions citing a case that does not exist, while saying plainly that AI used for research is welcome as long as a person verifies what it produces. The Supreme Court has also circulated draft rules for AI in courtrooms, built on a simple principle: the judge decides, and AI only assists.

So the bench has moved. The gap is everywhere else, and this is where India actually has to choose.

Start with data: India has its data-protection law, the DPDP Act, and the government notified the rules in November 2025. The real obligations do not bite until May 2027. India also picked its own approach to sending data abroad. Under the rules, personal data can leave the country unless the government names a place it may not go, and so far it has named none. So whether a client’s file can sit on a US company’s model is not yet fixed by law. It is a choice India is still holding in its own hands.

Then privilege: If a lawyer pastes a client’s confidential file into a public AI tool, has privilege been waived? No Indian court has answered that. Privilege now sits in the Bharatiya Sakshya Adhiniyam. The safer reading is that a public consumer tool looks a lot like handing the file to an outsider, while an enterprise system under a contract that bars the vendor from training on your data stands on firmer ground. That distinction is exactly what a firm like Freshfields paid for when it chose a closed deployment. Indian firms will have to make the same call, and today they make it with no rule to point to.

And the Bar Council has not yet written the rulebook the Supreme Court asked for. That is the real work left. The courts have said what happens when AI goes wrong in a filing. Nobody has yet said, in plain terms, what a competent Indian lawyer is expected to do with these tools before a matter ever reaches a judge: when to disclose that AI helped draft a filing, and what counts as properly checking what it wrote.

India does not need to wait for anyone else to settle this. It has a data law of its own and courts that have already acted. These rules are better written from India’s own courtrooms and its own data law than inherited from someone else’s. The technology is ready. Freshfields showed the scale. The Supreme Court has shown it will act. What is missing sits with the Bar Council and the courts, and with every firm deciding this week whether to let a model near a client file. That is a decision worth making on purpose, and making soon.