What AI does well for lawyers — and where the stakes punish carelessness

The honest taxonomy

Generative AI — the GPT-class chatbots, the Claude-style assistants, the copilots embedded inside Microsoft Word and Google Docs — is reliably useful for a narrower band of legal work than the brochures suggest. The band still covers a meaningful share of what a Nepali advocate, in-house counsel, or junior associate does on an ordinary Tuesday. But it has edges, and the cost of walking past those edges is higher in law than in any neighbouring profession. An accountant who quietly miscites an IRD circular can usually fix it before filing. A lawyer who quietly cites a non-existent Supreme Court decision in a writ petition has, in one stroke, deceived the court, breached the duty of competence owed to the Bar Council, and exposed the client to a possibly fatal credibility loss.

The clearest way to sort the work is to ask one question of every task: how cheaply can I verify the output, and how catastrophic is a quiet error? If verification is fast and the cost of a silent mistake is small, AI is a force-multiplier. If verification is slow and the cost of a silent mistake is large — a lost case, a sanction from the bench, a malpractice claim, a reference to the Bar Council’s disciplinary committee — then AI is a liability dressed as a shortcut.

What AI does well

Six categories, in roughly descending order of how confident you can be.

1. Structured first-draft summaries from material you provide. Hand the model a stack of intake notes, a witness statement, or a long contract, and ask for a structured summary — a chronology, a clause map, an issue list, a parties-and-obligations table. The model arranges what you gave it. It does not invent facts that were not in the source. A senior reviewer who has read the same material can sanity-check the structure in minutes, where writing it from scratch would take an hour. For a fifty-page loan agreement under the Banks and Financial Institutions Act, a clause map — “definitions, conditions precedent, representations, covenants, events of default, governing law” — is exactly the kind of dry skeleton the model produces well.

2. First drafts of routine letters and templates. Notice letters under section 95 of the Contract Act, demand letters for unpaid invoices, simple lease drafts, employment contracts that follow the Labour Act 2074, no-objection letters, vakalatnama covers — anything that lives ninety percent inside a known template — is fast to generate and easy to check. Keep your firm’s house-style versions in a folder, paste them as examples, and the model will produce a serviceable draft you adapt rather than retype.

3. Plain-Nepali translation of legal jargon for clients. A garments exporter in Birgunj does not want to hear about “indemnification, severability, and consequential damages.” She wants to hear, in plain Nepali, what she is promising and what could go wrong. The model handles this translation well, particularly when you give it the English clause and ask for client-facing Nepali at, say, an SLC reading level. Read it before sending — Nepali legal nuance is uneven in the model’s output — but the time saved on every client call is real.

4. Candidate arguments and counter-arguments. Before you draft your written submission, ask the model to list the strongest arguments your opponent is likely to raise and the strongest answers to each. Treat the output as a checklist, not a script. You will reject half. The remaining half will catch the angle you would otherwise have noticed at the rejoinder stage, after you had already committed to a weaker line.

5. Structured proofreading and consistency checks. Defined terms used before they are defined, party names that drift between “the Borrower” and “the Customer,” cross-references to clauses that no longer exist after edits, dates that contradict each other across the recitals and the schedules — these are the bugs that cost hours to find by eye. A model trained on long documents catches most of them in one pass. Treat its output as a list of candidates to check, not a list of edits to accept.

6. Translating between Nepali and English drafts. A petition drafted in Nepali for the District Court and a covering opinion in English for a foreign client are two documents the same firm often needs in one matter. The model handles the back-and-forth competently for prose; for the formal phrasing of prayers and reliefs, it produces a starting point you will polish by hand.

What AI does badly, or unsafely

The same technology fails — and fails most dangerously when it looks like it is succeeding — on a different cluster of tasks. These are the ones that have already produced sanctioned lawyers in other jurisdictions and will produce them here.

1. Quoting specific provisions of Nepali Acts. The model has read about the Civil Code 2074, the Criminal Code 2074, the Companies Act 2063, the Banks and Financial Institutions Act, and the Labour Act 2074. It has not necessarily read the current text after the latest amendment, and it has very rarely read the official Nepal Gazette. It will confidently produce a section number that does not exist, attribute a provision to the wrong chapter, or paraphrase a 2068 version of a clause that was amended in 2076. For any submission, opinion, or advice that turns on a specific provision, verify against the bare Act on the Nepal Law Commission website or a current commercial commentary. Treat any citation the model gives as a lead, not a source.

2. Citing case law from the Supreme Court of Nepal. This is the single most dangerous category. Models trained largely on English-language legal corpora have read very little Nepal Kanoon Patrika. When pressed for Nepali case authority, they will produce a citation in the right format — case name, NKP volume, year, page — that is entirely fabricated. The reasoning sounds correct. The judges named are real. The proposition is plausible. The case does not exist. Every citation to NKP, to a High Court decision, or to a District Court order must be opened in the Supreme Court’s online repository or the bound NKP before it goes into any document with your signature on it.

3. Reading dense regulatory text as final authority. A Nepal Rastra Bank unified directive, an Insurance Authority circular, a SEBON guideline — these are dense, frequently amended, and contain the operative language clients pay you to read. The model can produce a useful skeleton of “what this document appears to say.” You can use that skeleton to find the paragraph you need to read yourself. You cannot use the skeleton as the answer to the client.

4. Deciding borderline legal positions. Whether a particular transfer triggers capital-gains liability under section 95Ka, whether a clause amounts to a penalty rather than liquidated damages, whether a director has crossed the line into oppression under the Companies Act — these are the questions the client is paying you, and not the chatbot, to answer. The model can list the considerations and the cases on each side. It cannot weigh them in your client’s facts, and the professional consequence of getting it wrong does not transfer to the software.

5. Signing pleadings, opinions, or submissions. The act of putting your name on a document is a representation to the court, to the client, and to the Bar Council that a competent lawyer has formed a professional view. The model can help you draft. It cannot form the view. Your signature is the load-bearing thing in the document, and nothing in your subscription tier changes that.

The verification test, applied to legal work

Put the two lists together and one rule falls out. Before you use AI on any legal task, ask: if this output is wrong in a way I do not immediately notice, what does it cost — and how long would it take me to catch the error from a source I already trust?

In accounting, the cost of a silent error is often a restated set of accounts and a missed filing deadline. Both are recoverable. In legal work, the cost of a silent error is often a lost case, a sanction from the bench, a complaint to the Bar Council, or a malpractice exposure that follows you for years. The asymmetry is the point. The same model, used on the same draft, is safe in one profession and dangerous in another, because the cost of a quiet failure is not the same. A Nepali advocate should set the verification bar correspondingly higher — not because the technology is worse here, but because the consequence of trusting it carelessly is.

This is not a counsel of paralysis. Most of what fills your week — drafting, summarising, translating, structuring, listing arguments, proofreading — clears the test comfortably. The professional act is knowing, for each hour of the day, which side of the line you are on, and never letting the model’s confidence substitute for your own verification.

Check your understanding

Quick check

—

A junior associate at a Kathmandu firm has been asked to produce a clause-by-clause summary of a fifty-page facility agreement that the partner will negotiate in the morning. Which use of AI is the safest fit?

1. AAsk the model to list the leading Supreme Court of Nepal decisions on lender remedies under the Banks and Financial Institutions Act, and cite them in the summary.
2. BPaste the facility agreement into the model and ask it for a structured clause map — definitions, conditions precedent, covenants, events of default — then verify against the document.
3. CAsk the model to decide whether the indemnity clause is enforceable under Nepali law and to put that conclusion in the summary for the partner.
4. DAsk the model to draft and sign the negotiation note on behalf of the partner.

Check answer→

Quick check

—

The single best question to ask before using AI on a specific legal task is —

1. AHas the Bar Council issued formal guidance on AI use yet?
2. BDoes my firm have a paid subscription, or only the free version of the tool?
3. CIf the output is silently wrong, how costly is the error to the client, the court, and my standing — and how quickly can I verify it against a trusted source?
4. DHas any senior advocate in the office tried the tool before?

Check answer→

What comes next

Knowing which categories of legal task AI handles well is half the picture. The other half is knowing where in your actual working week those categories land — which stages of intake, contract review, drafting, research, court appearance, and billing change, and which do not. The next section walks through a typical week for a Nepali advocate and marks the map.