Chapter 02 · Section I · 17 min read
Reviewing a draft someone sent you
The single highest-leverage use of AI in a Nepali law office is the first-pass read of a contract someone else drafted — done as a structured workflow, never as a shortcut around reading.
The most common thing a Nepali advocate or in-house counsel does in a week is read a contract that someone else wrote. A vendor agreement from a Singapore counterparty, a distribution contract from an Indian principal, a fat shareholders’ agreement from a Mauritius investor, a fifteen-page MoU drafted by a deal team in London who have never heard of Nepal Rastra Bank. The draft arrives, usually late, usually long, usually one-sided, and the question is always the same: where are the landmines, what is missing, and what do we push back on. This is the work AI has, finally, made faster — provided you treat the model as a fast junior clerk and not as a substitute for opening the document and reading it yourself.
The structured first-pass workflow
The trap with a long imported draft is the temptation to read it linearly from page one, run out of attention by clause 14, and miss the change-of-control bomb on page 31. A model fixes this — not by replacing the read, but by giving you a map before you start. The workflow that actually works in a Nepali firm has four steps, in order.
1. Redact and load. If you are using a public-tier tool (the consumer ChatGPT, the consumer Claude, a free Gemini account) you must strip party names, project codenames, sums above a threshold, and anything that identifies the deal before pasting. If your firm has a paid enterprise tier with a no-training contract, this matters less but still matters for client-confidence reasons. The discipline is the same either way: never paste a draft you have not first looked at for names.
2. Ask for a clause-by-clause map, with risk flags. The model’s strongest move is structural. Ask it to walk the contract from clause 1 to the end, give each clause a one-line summary, and flag any of the standard risk terms — indemnities (scope, caps, carve-outs), liability caps (especially exclusions for gross negligence), IP assignment (work-for-hire vs licence-back), governing-law and jurisdiction surprises, termination triggers, change-of-control, exclusivity, MFN, non-compete, audit rights, data-processing obligations. The output is a one-page table you can read in three minutes that tells you where to focus.
3. Ask for a missing-clause check. This is the move most lawyers under-use. The model has seen tens of thousands of contracts of every kind; it knows what should be in a distribution agreement or an SaaS contract or a shareholders’ agreement and is excellent at noticing the absence. No force majeure clause? No notice provision specifying address and mode of delivery? No boilerplate severability or no-waiver? No dispute-resolution clause at all (you would be surprised how often)? The model will list these, and you will catch in three minutes what would have taken a careful linear read to surface.
4. Ask for a proposed redline list, ranked. Not the redline itself yet — a list, ranked from most to least important, of the changes you would push back on if you were representing the receiving party. This gives you and the partner something to triage against before any real drafting starts.
A prompt pattern you can paste
The shape matters more than the model. Adapt this to the document type:
You are assisting a Nepali advocate reviewing a draft contract on behalf of the [buyer / licensee / investee / recipient party]. The draft is pasted below. Produce four outputs, in order. (1) A clause-by-clause map: one row per clause, with the clause number, a one-line summary, and a risk flag (none / low / medium / high) with a brief reason where the flag is medium or higher. (2) A missing-clause check: list any standard clauses you would expect in a contract of this type that are absent or only weakly addressed. (3) A ranked list of the top redlines you would propose, with a one-sentence rationale for each. (4) Any clauses that look like they were drafted under a non-Nepal legal frame (US / UK / Indian / Singaporean) and may need localisation. Do not quote clause wording verbatim in the summaries — refer by clause number so I will open the document.
That last sentence is deliberate. It forces you to open the contract.
The non-negotiable verify step
This sounds obvious and it is the rule most often quietly broken. The model summarises a liability cap as “capped at fees paid in the preceding 12 months” — but the actual clause says “fees paid in the preceding 12 months, excluding any amounts in respect of professional services,” and your client’s deal is 80% professional services. The model summarises an indemnity as “mutual” — but the actual clause has a definition of “Indemnified Party” that quietly excludes affiliates, and your client’s group operates through three Nepali subsidiaries. These are not exotic failures. They are the normal failure mode of a tool that compresses language for speed.
The discipline is: the model’s map tells you where to read; you still read.
Common LLM failures on contract review
A few patterns recur often enough in Nepali practice that they are worth pre-loading into your attention.
Cross-references between clauses get missed. The contract defines “Material Breach” in clause 1.1, uses it in the termination clause at 14.3, and then quietly modifies the threshold in a schedule at the back. The model sees each in isolation and reports each correctly; the interaction — that the schedule guts the termination right — slips past. Always ask explicitly whether any schedule, annexure, or side letter modifies a main-text clause.
Defined terms used inconsistently. A draft will define “Confidential Information” with a carve-out for publicly known material, then use “confidential information” (lower case) in three later clauses without the carve-out. A careful human reader catches this; a model summary almost never does.
Defaulting to US/UK frames on jurisdiction. Ask a model what is unusual about a governing-law clause and it will compare against a Delaware or English baseline. A clause specifying Nepali governing law and arbitration in Kathmandu under NEPCA Rules will get flagged as “non-standard” — when in fact it is exactly what the deal needs. Tell the model the parties are Nepali and the deal is Nepali, and the noise drops.
Liability cap carve-outs read wrong. The model will report “limited liability with cap” and move on. The actual clause may exclude gross negligence, wilful misconduct, IP indemnity, and confidentiality breach from the cap — which is normal — but it may also exclude “any payment obligation under this Agreement,” which silently turns the cap into nothing. Always ask the model to list the cap’s exclusions explicitly.
The Nepal-specific watch-list
Every contract review for a Nepal-touching deal should run through a short local checklist. These are the things foreign drafters routinely get wrong and that local counsel exists to catch.
1. Governing law. Is Nepali law specified? If foreign law is proposed for a deal where the Nepali party will perform in Nepal, is that actually enforceable, and is it commercially what the client wants?
2. Arbitration seat and rules. A KTM-seated arbitration under NEPCA Rules is the local default and is enforceable. ICC and SIAC are workable for cross-border deals but bring cost and the question of enforcement of the award in Nepal under the New York Convention (Nepal is a party). A clause that names “arbitration in London under LCIA Rules” for a domestic Nepali deal is almost always wrong.
3. Language clause. If the contract will be filed with a Nepali authority — Office of the Company Registrar, IRD, a sector regulator — there must be a Nepali version, and the contract should say which version controls. The model will not flag this; you must.
4. Stamp duty. Who pays, when, and at what rate. For most commercial agreements this is small money but it is an enforceability question if missed.
5. FX repatriation and NRB compliance. Any cross-border payment flow — royalties, management fees, dividends, loan principal and interest — must be checked against the prevailing NRB circular regime. The model does not know the current circulars. You do, or you must check before you sign off.
Check your understanding
Quick check
—A 40-page distribution agreement arrives from a Mumbai-based principal at 6pm and the partner wants a clause map by morning. Which workflow is safest and most time-efficient?
Quick check
—You ask the model to flag unusual clauses in a draft contract between two Nepali companies. The model flags the governing-law clause (Nepal) and arbitration seat (Kathmandu under NEPCA Rules) as non-standard and recommends Delaware law and SIAC arbitration. What is the right move?
What comes next
Reviewing an imported draft is the receiving side of contract work. The drafting side — producing the document in the first place, either from your firm’s precedent or from scratch — has its own discipline, its own prompt patterns, and its own list of things a model will quietly leave out. The next section is on drafting.