On 2 July 2026 the Supreme Court set aside NCLT and NCLAT orders in an insolvency matter because the reasoning rested on judgments that do not exist. It declared zero tolerance for unverified AI-generated authority and held that an advocate who cites it can be guilty of misconduct.
It did not tell lawyers to stop using AI. It told them where the responsibility sits. AI can produce a usable first draft of a notice, plaint or agreement in minutes. It cannot vouch for anything that draft asserts.
This guide covers what to hand over to AI drafting, what to check before a draft is filed or sent, and how to judge a tool before client work goes through it.
What is AI drafting for lawyers?
AI drafting for lawyers means giving a language model the facts, your instructions and your templates, and getting back a first version of a notice, plaint, affidavit, application or agreement to work on. The advocate reviews, corrects and signs. The draft is a starting point, never an authority.
It is strongest at structure and repetition: recitals, standard clauses, formatting, reworking a precedent for new parties. It is weakest wherever the document depends on something outside the facts you gave it, such as a citation, a limitation period or a current section number.
For the wider view, read how AI is changing legal practice in India.
What the 2 July 2026 ruling changed?
The ruling raised the price of an unchecked citation. It did not make AI-assisted drafting improper.
In Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668), a bench of Justices P.S. Narasimha and Alok Aradhe set aside the NCLT and NCLAT orders. Of the six judgments those orders cited, three did not exist and the other three carried paragraphs nobody wrote (iPleaders). A decision resting on even a trace of hallucinated material, the Court held, has no standing in law. It asked the Bar Council of India to constitute a committee to frame guiding principles and disciplinary consequences for advocates who put such material before a court (Free Press Journal).
The BCI had set out its own line a month earlier, in a reply to Parliament. AI use is not prohibited, and the Council lists drafting assistance among the permitted uses. But every statute, precedent and quotation that AI produces must be independently verified, and an advocate cannot blame a machine for a fake citation. Disciplinary action can follow under Sections 35 and 36 of the Advocates Act, 1961 (Bar & Bench).
The working rule: draft with AI, verify everything it asserts.
Case law is missing from the middle column on purpose. If a draft cites a case you did not supply, treat it as unverified until you have read the judgment. AI drafts can also carry IPC or CrPC numbers into a matter governed by the newer codes, so confirm the code first.
A worked example: a Section 138 notice
Take a bounced cheque. You give the tool the payee and drawer, the cheque number and amount, the date it was presented, the date and reason on the bank's return memo, and a 15-day demand. You ask for a legal notice in your firm's format, with placeholders for anything missing.
The first version is usually sound on structure: numbered facts, the dishonour, the demand, the consequence of default. The checks are where drafts slip quietly:
- The 15 days run from receipt of the notice, not from the day you send it. Counting from dispatch is a common slip.
- The notice must go out within 30 days of the bank's dishonour information reaching the payee. Check the return memo date against your dispatch date.
- Cheque number, date and amount must match the return memo exactly.
- Name the drawer as it appears on the account. A trading name or the wrong signatory can cause trouble later.
Ten minutes of checking turns the draft into a notice you can sign. The saving is in the typing and formatting. The protection is in the checking.
A drafting workflow that holds up
A good drafting session is a loop you control from the first prompt to the export.
- Start from your own precedent. Feed the tool your saved template for the document type, so structure, house language and court conventions come from you.
- Give facts as inputs, not as a story. Parties, dates, amounts, forum, relief sought.
- Say what to leave out. No case law, and a placeholder for anything it cannot know. A good tool leaves the blank rather than guessing.
- Revise by instruction. Ask for a clause, a tighter prayer, a firmer tone. Each pass should change the draft, never your facts.
- Verify, then save the draft against the matter.
A good instruction names the document, the parties, the forum, the dates and amounts, the relief and the tone, and says what to leave out. For example:
Draft a legal notice under Section 138 of the Negotiable Instruments Act for the payee, [name], against the drawer, [name]. Cheque no. [number] dated [date] for ₹[amount], returned unpaid on [date] for insufficient funds. Demand payment within 15 days of receipt. Formal tone, numbered paragraphs. Do not cite case law. Leave a placeholder for anything not stated above.
Specific inputs make a draft checkable: every number in the output should trace back to a number you supplied.
The demo on GYL's AI legal drafting software page shows the loop: asked to draft an eviction notice for a landlord against a housing board, Vidhilekh AI builds it on the saved Eviction Notice template. A follow-up asking for an arrears clause adds it as a new paragraph and leaves the amount blank for the advocate to confirm.
Where AI drafts go wrong
AI drafts fail in predictable ways, and each failure maps to a check.
- Invented authority. Citations that look right and do not exist. This is what the Supreme Court acted on.
- Stale law. A model's knowledge stops at its training date, so it can miss a recent amendment or notification.
- The wrong code. IPC or CrPC numbers where BNS or BNSS applies.
- Confident gaps. A missing date or amount gets filled with something plausible instead of a question.
- Unrequested edits. Ask for one change, then confirm the rest of the draft is untouched.
None of these is obvious at a glance, which is why the verification pass is not optional.
The verification pass
Whatever the tool wrote, your signature is on it. Run every draft through these checks before it is filed or sent.
- Authorities. Open each cited judgment in an authorised source such as SCC Online, Manupatra or the official reports. Confirm that it exists and that the paragraph you rely on says what the draft says it does.
- Statutes. Check section numbers and the code they sit in, and look for amendments made after the model's training data ended.
- Dates and limitation. Work them out yourself from the source documents.
- Facts. Match every name, amount, address and date to the client's papers.
- Forum. Court, bench, jurisdiction, valuation.
- Anything you did not supply. A case, figure or proposition the model added is unverified until you have confirmed it.
Keep a short note of what the tool produced and what you changed. If a client or a court asks how a draft was prepared, you can answer in a sentence.
How to choose an AI drafting tool
Judge a tool by what it does when it does not know.
- Your templates first. Can it draft from your saved precedents, or only from its own?
- Indian document types. Ask for a Section 138 notice, a bail application and a written statement. Look at the format, the party descriptions and the prayer.
- Placeholders over guesses. It should leave a blank for an amount or date you did not give, and it should not volunteer authorities you did not ask for.
- Data handling. Ask whether your inputs are retained or used to train shared models, and how one matter is kept apart from another.
- Workflow fit. Drafts that live against the case, beside the hearing dates and the client record, save retyping and misfiling.
Then run a twenty-minute test. Take a document you filed last year, strip the identifiers, and ask the tool for the same document from the same facts. Compare it with yours line by line. The gaps will tell you more than any demo.
Who gets the most from AI drafting
The gain differs by practice.
- Solo advocates often draft the same notices, replies, affidavits and applications every week. Templates plus AI cut the retyping and leave time for the file.
- Law firms gain consistency: one house style across juniors, and partner review that goes to facts and law instead of formatting. For the wider workflow, see the benefits of legal case management software for Indian firms.
- Enterprise legal teams can use first drafts for routine contracts and vendor paper, and keep counsel's time for the points that need negotiating.
Confidentiality and your duties
AI does not change what you owe a client. It changes where their information can end up.
Advocates are bound by confidentiality under the Bar Council of India Rules, and professional communications are privileged under Section 132 of the Bharatiya Sakshya Adhiniyam, 2023 (iPleaders). The BCI has told Parliament that an advocate cannot put privileged communications, confidential instructions or sensitive strategy on an insecure platform (Bar & Bench).
In practice: anonymise facts the draft does not need, and read a tool's data terms before a matter goes through it. The Digital Personal Data Protection Act, 2023 adds another layer. Its Rules were notified in November 2025, and the main obligations on data fiduciaries apply from May 2027 (DPDP key dates).
