What AI can and cannot do in Indian legal drafting

There is a great deal of noise about AI and legal work at the moment, most of it either breathless or dismissive. Having built drafting tools for Indian litigation and watched advocates use them daily, we have a fairly specific view of where the line sits.

Where it genuinely helps

The honest answer is that AI is good at the parts of drafting that are structural rather than judgemental.

• Getting to a first draft. A blank page is expensive. A structured draft with the right headings, numbered paragraphs, the correct prayer format, and the statutory provisions cited in the right places removes the worst part of the task, even when half of it gets rewritten.
• Summarising a long record. Reading forty pages of orders to establish the procedural history of a matter is work a model does quickly and reliably, because it is extraction rather than reasoning.
• Finding the case you half-remember. Searching a judgment corpus by the substance of a proposition, rather than by keyword, is where semantic search genuinely outperforms the alternative.
• Consistency checks. Catching that a party is named two different ways across a filing, or that a date in paragraph 12 contradicts paragraph 40, is tedious for a person and trivial for a machine.

Where it fails, and fails confidently

The failures matter more than the successes, because they do not announce themselves.

A general-purpose model asked for authority on an Indian proposition will, with some regularity, produce a citation that is formatted perfectly and does not exist. This is not a bug that will be patched away — it is a consequence of how the models work. The mitigation is architectural: the model must retrieve from a real corpus and cite what it retrieved, rather than generating citations from memory. Every case law reference Lexshastra surfaces is a document in our own corpus of Indian judgments, retrievable and readable in full. If it cannot find one, it says so.

A model that cannot find authority should say nothing. The failure mode to design against is not silence, it is confident invention.

The second failure is judgement. A model can tell you what the law says. It cannot tell you whether to raise a point that will annoy a particular judge, whether your client will actually turn up, or whether this is the matter to settle. Those are the decisions that make an advocate worth their fee, and nothing in the current technology comes close to them.

The third is local practice. Filing conventions vary by court, by registry, and sometimes by clerk. A model trained on text has no access to the fact that a particular registry rejects a particular annexure format. That knowledge lives in the chamber, not in the corpus.

Why the advocate stays in the loop

This is a design principle for us rather than a disclaimer. Lexshastra's drafting tools produce drafts into a document you own and edit. The AI suggests in a sidebar; it does not reach into your text and change it. Nothing is filed, sent, or served by a machine. Every output is a draft for an advocate to review, correct, and sign.

Partly that is professional responsibility — the signature on a filing is a person's, and so is the accountability. But it is also simply what produces better work. The useful framing is not automation but a capable junior: fast, tireless, well-read, occasionally wrong in ways that need catching, and never the one who decides.