Most public discussion of legal AI in India begins and ends with the courts. The reasons are obvious: tens of millions of pending cases, a national digitisation programme in eCourts, and the visible question of whether technology can shorten the time it takes to get justice. But the courts are only one consumer of legal work. Behind every filing sits a law firm or a corporate legal team that researched the law, drafted the documents, reviewed the contracts and advised the client. That is where most legal labour is spent, and where AI is being adopted fastest, often without much policy.
This piece looks at legal AI in India from the side of the profession rather than the bench. It reviews what is known about adoption among lawyers, sets out the rules that already govern how Indian advocates may use AI, even though none were written with AI in mind, and offers an original task map that classifies legal work by how suitable it is for automation and how much harm an error could cause. The map is intended for managing partners, general counsel and legal operations leads who must decide where AI belongs in their practice and where it does not.
The argument is straightforward. The question for Indian legal teams is no longer whether lawyers will use AI tools, because many already do. It is whether they will use them inside a framework of verification, confidentiality and accountability that protects clients and the court, or outside one.
The scale of legal work in India
The judiciary's numbers set the context. The eCourts Phase III project, approved by the Union Cabinet in September 2023 with an outlay of ₹7,210 crore, had computerised 18,735 courts and placed 27.64 crore orders and judgments on the National Judicial Data Grid by December 2024, with 3.38 crore cases heard virtually. Pendency, according to a Lok Sabha reply of February 2026, stood at 4.77 crore cases in district courts, 63.66 lakh in High Courts and 92,101 in the Supreme Court.
Each pending case represents legal work for at least two sets of lawyers, and litigation is a fraction of what the profession does. Contract drafting and review, regulatory compliance, due diligence for transactions, employment matters, intellectual property and dispute avoidance occupy corporate legal teams and transactional practices. The digitisation of the courts also changes this work: when judgments and orders are searchable at national scale, legal research becomes a data problem, and the lawyers who can query that data efficiently gain an advantage.
What adoption looks like among lawyers
India-specific survey data on lawyers' use of AI is thin, so global surveys are the best available guide, with the caveat that they are dominated by respondents from North America and Europe.
Thomson Reuters' 2025 Generative AI in Professional Services report, based on nearly 1,800 professionals across legal, tax, accounting, risk and government, found that legal use of generative AI nearly doubled, from 14% in 2024 to 26% in 2025. Among law firm respondents, 89% believed AI could be applied to their work, yet 48% lacked formal policies on generative AI. A Thomson Reuters analysis of the same research identified six use cases cited by half or more of legal respondents: document review, document summarisation, legal research, brief or memo drafting, contract drafting and correspondence drafting.
The company's 2025 Future of Professionals report, published in June 2025 and based on 2,275 professionals, found that only 22% of organisations had a visible AI strategy, and that those with one were twice as likely to report AI-driven revenue growth as organisations with informal or ad hoc approaches. Respondents expected AI to save them five hours a week within a year.
Two readings follow for India. First, the gap between belief in AI's usefulness and the existence of policy is likely to be at least as wide in Indian firms as in the global sample, given the absence of profession-specific guidance discussed below. Second, the six leading use cases are all tasks where a lawyer remains responsible for the output, which makes verification, not access, the binding constraint.
The rules that already apply
There is no AI-specific rule for Indian advocates. That does not mean there are no rules. A column in Bar and Bench by Urvashi Misra and Shailja Rawal of AZB and Partners, published in February 2026, notes that the Advocates Act, 1961 and the Bar Council of India Rules set out duties that bind any advocate using any tool: not to mislead the court, to act in the client's best interests, to maintain confidentiality and to give competent advice. The authors describe the absence of AI-specific guidance as a concerning lacuna and call for a pan-India framework.
The Bar Council of India has since stated its position. Bar and Bench reported in June 2026 that the BCI told Parliament that AI use is not prohibited, that AI-generated research must be independently verified before it is relied on in court, and that advocates who file false cases, fake quotations or misleading propositions remain accountable and may face disciplinary proceedings under Sections 35 and 36 of the Advocates Act. In the BCI's words as reported: "The machine neither signs the pleading nor owes duties to the court. The advocate does." The BCI also said it has sufficient power under the Act to issue clarifications and professional standards for AI-assisted work when required.
The judiciary has moved faster on its own use. The Kerala High Court's policy for the district judiciary, reported by MediaNama in July 2025, prohibits judges from using AI tools to arrive at any findings, reliefs, orders or judgments, restricts the use of public cloud-based generative AI tools because uploading case facts or privileged material may breach confidentiality, permits only approved tools for official work, and requires that all outputs including citations be meticulously verified, that staff be trained, and that courts maintain an audit of every instance of AI use. The AZB column notes that the Bombay Bar Association has issued similar guidelines and that a Supreme Court white paper of November 2025 on AI and the judiciary emphasised human verification, confidentiality and disclosure. The risk these rules address is not hypothetical. In Gummadi Usha Rani v. Sure Mallikarjuna Rao, decided on 27 February 2026 and reported by SCC Online, the Supreme Court considered a trial court order that had relied on precedents the defendants said were fabricated, and which the Andhra Pradesh High Court had acknowledged appeared to be AI-generated. The Supreme Court issued notice to the Attorney General, the Solicitor General and the Bar Council of India, appointed an amicus curiae, and observed that "a decision based on such non-existent and fake alleged judgments is not an error in the decision making", indicating that such reliance could amount to misconduct.
For corporate legal teams, a further layer applies. Contracts, disputes and investigations involve personal data, and the Digital Personal Data Protection Rules notified in November 2025 begin an 18- month phased implementation of the DPDP Act 2023. Sending client documents to a public AI service is a data processing decision, not just a productivity choice.
A task map for legal AI in India
The original contribution of this piece is a map of legal tasks along two dimensions. Automation suitability asks how well current AI tools perform the task: whether inputs are mostly text, whether a correct output can be checked efficiently, and whether the task is repeated at volume. Error risk asks how much harm an undetected mistake could cause to the client, the court or the lawyer's standing, and how visible that mistake would be before it caused harm.
Crossing the two produces four zones:
Zone A, automate with sampling: high suitability, lower risk. AI performs the task at volume; humans review a sample and exceptions. Zone B, assist with mandatory review: high suitability, higher risk. AI produces a draft or analysis; a qualified lawyer reviews every output before use. Zone C, human-led with AI support: lower suitability, variable risk. The lawyer leads; AI helps with search, summarisation or formatting. Zone D, do not delegate: tasks that require professional judgement, accountability or presence that cannot be transferred.
| Task | Suitability | Error risk | Zone | Control required |
| Translating documents and judgments between English and Indian languages for internal understanding | High | Low to medium | A | Sample review; certified translation where it will be filed |
| Summarising long judgments, orders or case files for internal briefing | High | Medium | A or B | Spot checks against source; never cite a summary |
| Classifying and extracting clauses across a contract portfolio | High | Medium | A | Sample review; exception queue for unusual clauses |
| Tracking new regulations, circulars and notifications against internal obligations | High | Medium | A | Human sign-off on what changes in practice |
| First-pass due diligence review of a data room | High | High | B | Lawyer review of every flagged and a sample of unflagged documents |
| Legal research and identifying relevant precedents | High | High | B | Every citation verified against an authoritative source before use |
| Drafting contracts from an approved playbook | High | Medium to high | B | Lawyer review of every draft; deviations from playbook escalated |
| Drafting pleadings, petitions, written submissions and legal opinions | Medium | Very high | B or C | Lawyer authorship and full verification; the advocate signs |
| Litigation docket and deadline management | Medium | High | C | Human-confirmed deadlines; AI as alert, not authority |
| Negotiation strategy and client advice | Low | High | C | Lawyer judgement; AI for preparation only |
| Court appearance, oral argument and cross-examination | Low | Very high | D | Not delegated |
| Final sign-off, certification, and professional responsibility | Not applicable | Very high | D | Not delegated |
The map makes several practical points visible.
The largest efficiency gains are in Zone A and Zone B, which together cover most of the six use cases identified in the Thomson Reuters research. But the two zones need very different controls. A firm that treats legal research as if it belonged in Zone A, reviewing only a sample, is exposing itself to exactly the fabricated-citation failure that reached the Supreme Court in 2026.
Translation is unusually valuable in India. With court records, client documents and regulatory material spread across English and many Indian languages, an AI that produces a reliable working translation for internal understanding removes a real bottleneck. Filing that translation is a different matter and moves the task into Zone B or requires certified translation.
Zone D is narrower than sceptics claim and wider than enthusiasts suggest. It is defined not by what AI cannot do technically but by what the profession's rules assign to a person: signing, appearing and taking responsibility.
Building the controls
The task map is only useful if the controls in the final column are real. Four practices make them so.
Approve tools, not just uses. The Kerala High Court's approach of permitting only approved tools translates directly to firms and legal departments. An approved list, with each tool assessed for where data is processed, whether inputs are retained or used for training, and how outputs can be traced to sources, is the foundation. Regulated clients may require that some tools run within the firm's or client's own environment.
Verify citations mechanically. For Zone B research tasks, every case, statute and quotation should be checked against an authoritative database before it enters a document. The check can itself be partly automated, by matching citations against judgment repositories, but the sign-off must be human.
Keep an audit trail. The Kerala policy's requirement to record each use of AI and the human process used to check it is good practice for any legal team. It protects the lawyer if work is challenged and generates the data needed to decide which tasks can safely move between zones.
Train for verification, not prompting. The scarce skill is not getting a tool to produce a draft; it is recognising when a confident draft is wrong. Training should focus on failure modes: invented citations, misread holdings, outdated law and confidently wrong summaries.
A separate PanScience Research piece covers evaluation before deployment, and a separate Cluster 4 piece covers enterprise RAG and its failure modes, both of which apply to legal research tools that retrieve and summarise case law.
Practice and capacity building
Two examples from the PanScience portfolio illustrate the court-facing and capacity-building ends of this landscape. PanScience reports that its venture Nyaay AI provides judicial workflow automation, with deployments in more than 14 Indian High Courts and three national supreme courts (company- reported). The lessons from court-facing work, particularly the insistence on human verification and audit, carry over to firms and in-house teams, because the same documents move between them.
On capacity, BITS Law School and PanScience Innovations launched the PALETTE Centre, Professional Advancement in Law through Executive Training and Technical Education, in May 2025, as reported by Bar and Bench. Former Chief Justice of India D.Y. Chandrachud was chief guest and delivered the presidential address, and Attorney General R. Venkataramani was guest of honour. The centre is described as a hub to build AI readiness across India's legal system, offering training in AI- assisted legal research, case classification, document review and courtroom automation for lawyers, judges, judicial staff and students. At the launch, Dr Chandrachud said: "Lawyers today must not only interpret statutes but also understand how AI systems arrive at their recommendations." The remark captures the point of the task map: understanding how a tool reaches its output is what allows a lawyer to judge which zone a task belongs in. PanScience Research discloses that PALETTE is a PanScience initiative.
What law firms and general counsel should do now
For managing partners and general counsel, a sequence that works: 1. Inventory current use. Many lawyers are already using public AI tools. Find out which, for what, and with what data. 2. Map tasks to zones. Use the task map, adapted to the practice, to decide where AI may be used and with what control. 3. Approve a short list of tools. Assess data handling, retention and traceability; prefer tools that cite their sources. 4. Write a one-page policy. Cover approved tools, prohibited data, mandatory verification by zone, disclosure to clients where appropriate, and audit logging. 5. Measure and adjust. Track time saved and errors caught by task, and move tasks between zones only on evidence.
Conclusion
Legal AI in India is arriving through the profession's side door: individual lawyers and in-house teams adopting tools for research, review and drafting faster than their organisations write policy. The rules that matter already exist in the Advocates Act, the Bar Council of India's stated position and the judiciary's own guidance, and they converge on one principle: the lawyer, not the machine, owes duties to the court and the client. The task map in this piece turns that principle into practice by assigning each kind of legal work a level of automation and a matching level of control. Firms and legal departments that adopt such a map now will capture the efficiency gains while avoiding the failures that are already reaching India's highest court.
Methodology and sources
This article draws on court and government data, survey research, reported regulatory positions and an original task map. Court data come from editor-verified PIB material on eCourts Phase III and a Lok Sabha reply on pendency of February 2026. Adoption figures come from Thomson Reuters' press release of 15 April 2025 on the 2025 Generative AI in Professional Services report (thomsonreuters.com/en/press-releases/2025/april/from-incubation-to-integration-generative-ai- adoption-nearly-doubles-as-professional-services-reach-crossroads), its legal use-case analysis of 13 May 2025 (legal.thomsonreuters.com/blog/generative-ai-for-legal-professionals-top-use-cases), and its Future of Professionals press release of 26 June 2025 (thomsonreuters.com/en/press- releases/2025/june/the-ai-adoption-reality-check-firms-with-ai-strategies-are-twice-as-likely-to-see-ai- driven-revenue-growth-those-without-risk-falling-behind). Regulatory positions come from Bar and Bench on the BCI's statement to Parliament, 4 June 2026 (barandbench.com/amp/story/news/can- lawyers-face-action-for-ai-generated-fake-citations-heres-what-bci-told-parliament), the AZB and Partners column in Bar and Bench, 20 February 2026 (barandbench.com/amp/story/columns/at-the- crossroads-ai-ethics-and-legal-practice), MediaNama on the Kerala High Court policy, 21 July 2025 (medianama.com/2025/07/223-kerala-hc-bars-ai-tools-chatgpt-deepseek), and SCC Online on the Supreme Court's February 2026 order (scconline.com/blog/?p=377986). The Supreme Court white paper and Bombay Bar Association guidelines are referenced as reported in the AZB column. PALETTE details come from Bar and Bench, 27 May 2025 (barandbench.com/Law-School/bits-law- school-and-panscience-innovations-unveil-palette-ai-legal-hub). Global surveys are not India-specific. The task map, its zone assignments and control requirements are the author's analysis. Nyaay AI details are company-reported by PanScience Innovations (panscience.xyz).