Chapter 02 · Section II · 16 min read
Drafting from scratch and from precedent
The fastest way to a good Nepali contract draft is almost never to ask the model for "a standard agreement" — it is to feed the model your own precedent, then make it do the boring work of adaptation.
There are two distinct drafting modes a Nepali lawyer should learn to run cleanly with AI, and they are not interchangeable. The first is drafting from a precedent your firm already trusts — adapting your last good services agreement, your last clean shareholders’ agreement, your last well-worked distribution contract for a new deal. The second is drafting from scratch, when no usable precedent exists or the deal is genuinely novel. Almost every drafting problem you will have this year is the first kind; almost every mistake you will see junior lawyers make with AI is treating the first kind as if it were the second. This section is about both, with the bias clearly toward precedent — because that is the bias that produces good contracts faster.
From precedent — the better way
The most underused move in AI-assisted drafting is the simplest one. Paste your firm’s preferred precedent — the services agreement, the SPA, the lease, the distribution contract you have already settled — into the model up front. List the parties for the new deal. List the specific commercial terms (price, term, scope, territory, exclusivity, key conditions). Ask the model to adapt the precedent to the new deal, preserving the structure, the defined terms, and the house drafting style, and to flag anywhere it is making a substantive choice you should review.
What you get back is a draft that already speaks your firm’s language. The clauses are in the order you expect. The definitions are the ones your partners have already negotiated. The boilerplate is the boilerplate you have already been comfortable signing for years. The review is faster — because you are reading a familiar document with specific changes — and the client is better served, because the contract is built on a foundation your firm already knows how to defend.
Compare this to the common mistake: asking the model for “a standard services agreement governed by Nepali law.” What you get is a competent generic draft assembled from the average of the model’s training corpus — which means American structure, American defined-term conventions, American clause ordering, with a Nepali governing-law clause grafted on the end. It is workable, but you will spend more time reshaping it into something your firm is comfortable with than you would have spent adapting your own precedent in the first place.
From scratch — when you really must
Sometimes there is no precedent. A new kind of deal, a new sector, a one-off arrangement — a co-development agreement for an agri-tech pilot, a data-sharing arrangement between a hospital and a research institute, a hybrid licence-and-services deal for a fintech integration. In these cases the model is genuinely useful, but the discipline shifts. You are not adapting a known-good document; you are assembling one, and you are responsible for the assembly being correct.
The workflow that works: list every clause you want, with the substantive terms for each, before asking for any drafting. Not “a co-development agreement” but a structured brief — parties, recitals (what is each side bringing), scope of the development, IP ownership of background material vs foreground material, allocation of costs, milestones and acceptance, term, termination, confidentiality, publicity, warranties, indemnities (and their carve-outs and caps), insurance, force majeure, notices, assignment, governing law, dispute resolution, language, counterparts and execution. Then ask the model to draft each clause to your specification. The model is a much better stenographer than it is an architect; the architecture has to come from you.
Always specify, up front, the deal-level defaults that the model will otherwise guess at:
1. Governing law — Nepal, unless the deal genuinely requires otherwise. 2. Forum — a Nepali district court of appropriate jurisdiction, or KTM-seated arbitration under NEPCA Rules. 3. Language — English for execution, with a Nepali translation for filing if required, and a clear statement of which controls. 4. Signing protocol — wet signature or electronic signature under the Electronic Transactions Act, number of counterparts.
If you do not specify these, the model will produce a draft with Delaware law, New York courts, English-only language, and a US-style “executed in counterparts” boilerplate that quietly does not address what your client needs. You will catch it on review, but the review will be longer.
Common omissions models leave you to catch
Even with a clear brief, models systematically forget or under-specify certain clauses in commercial drafting. Build a personal checklist and run it on every draft — yours or the model’s. The recurring gaps in Nepali practice:
Indemnities scoped too narrowly. The model will draft a third-party IP indemnity and stop. The deal may also need an indemnity for breach of warranties, breach of confidentiality, regulatory non-compliance (data protection, AML), and personal injury / property damage where physical work is involved. The carve-outs matter too — the standard exclusions for the indemnified party’s own negligence are often left out.
Force majeure missing or boilerplate-only. A model will produce “fire, flood, war, act of God” and call it done. For Nepal, the clause needs to be specific enough to cover the things that actually disrupt performance here — bandh, prolonged power outages, fuel crisis, monsoon road closures, earthquake. Otherwise you spend the next dispute arguing whether a three-week bandh is “civil disturbance” within the meaning of clause 14.
Notice-and-cure period missing on termination. A bare “either party may terminate on material breach” is a litigation invitation. The clause needs a notice mechanism, a cure period (typically 30 days for monetary, 60 for non-monetary), and the consequence of failure to cure.
Assignment provision absent. Whether either party can assign, novate, or sub-contract — and whether change of control counts as an assignment — should be addressed. Models often draft this as a one-line “neither party may assign without consent” which is too thin for a serious deal.
Survival clause forgotten. Which clauses survive termination — confidentiality, indemnities, limitation of liability, dispute resolution, governing law — needs to be stated. Models leave this out maybe half the time.
Nepal-specific defaults to bake in
Beyond the clause checklist, certain Nepal-specific elements should be part of the model’s instructions from the first prompt, not added later:
1. A Nepali language clause if the deal will be filed locally — with Office of the Company Registrar, with IRD for VAT-related agreements, with a sector regulator (NTA, NRB, Insurance Authority).
2. Stamp duty allocation — typically borne by the party who benefits from registration, but should be stated explicitly to avoid the small but irritating last-minute dispute.
3. FX clauses for cross-border deals — payment currency, conversion mechanism (NRB reference rate on the payment date is the standard), responsibility for NRB approvals, treatment of FX losses.
4. Arbitration with NEPCA Rules where the parties want a Nepali-seat arbitration — and a fallback to ICC or SIAC only where the cross-border nature genuinely warrants it.
5. A clear PAN/VAT recital — both parties’ PAN, VAT registration status, and any applicable TDS withholding obligations on payments under the contract. This avoids the routine post-execution scramble at the first invoice.
A worked example — a five-year vendor services agreement
Take a concrete case. A Kathmandu IT services firm is providing managed IT services to a hotel chain headquartered in Pokhara, for five years, with monthly fees, an SLA, and an exclusivity provision in favour of the IT firm for the hotel’s IT outsourcing. A precedent-driven prompt looks roughly like this:
Pasted below is our firm’s standard managed services agreement template, which we have used successfully for similar engagements. Adapt it for a new engagement with the following parameters. Service provider: [Provider Name], a Nepali company with PAN [X]. Customer: [Customer Name], a Nepali company with PAN [Y]. Services: managed IT services covering helpdesk, network, server, and security operations across the customer’s six properties in Nepal. Term: 5 years from the effective date, with auto-renewal for 1-year terms unless either party gives 180 days’ notice. Fee: monthly fixed fee of NPR [Z], with annual escalation tied to year-on-year change in the NRB national CPI. Exclusivity: the customer will not retain any other managed IT services provider during the term, subject to the carve-outs in Annex A. SLA: as per the template, with the response-time matrix in Annex B. Governing law: Nepal. Dispute resolution: KTM-seated arbitration under NEPCA Rules, three arbitrators for disputes above NPR 5 crore, sole arbitrator otherwise. Language: English; Nepali translation only if required for filing. Produce the adapted draft, preserving the template’s structure and defined terms. Flag any clause where you made a substantive choice I should review, and any place where the template’s existing wording may not fit the new commercial terms.
The model returns a draft that is 80% your template, 20% the new deal — exactly the ratio that lets you review it intelligently in an hour rather than re-drafting it in a day.
Check your understanding
Quick check
—Your firm has used essentially the same managed services agreement template for fifteen engagements over the past four years. A new managed services deal lands. What is the most efficient and reliable way to produce the first draft using AI?
Quick check
—You ask the model to draft a vendor services agreement between two Nepali companies. The model produces a competent draft, but the governing-law clause specifies Delaware law and the dispute clause specifies New York state courts. What is the right correction?
What comes next
A draft, however carefully built, is the start of the conversation, not the end. The opposing side will respond with their own draft, their own redlines, their own counter-arguments. The next section is on using AI in the negotiation phase — generating arguments, drafting redlines, and knowing where the model’s contribution stops and the lawyer’s judgment about the relationship begins.