Negotiation playbooks and the redline conversation · AI for Lawyers

By the time a contract reaches the negotiation phase, the easy work is done. The boilerplate has been agreed, the obvious clauses have settled, and what remains is the small number of provisions — typically the liability cap, the indemnity scope, the exclusivity, the IP ownership question, the termination triggers — where the two sides actually disagree about money or control. This is where AI’s contribution shifts in character. In review and drafting the model is a fast clerk; in negotiation it becomes something closer to a sparring partner — generating arguments you would not have thought of, predicting counters you will face, and producing redline language on demand. But it remains a sparring partner who does not know the room, the relationship, or what your client can afford to lose, and that limit shapes everything in this section.

The most reliable use of AI in negotiation is what economists would call candidate generation. You have a contested term — say a liability cap pegged at one year of fees with no exclusions for gross negligence or IP indemnity. You ask the model for a structured menu: arguments in favour of pushing for change, likely counter-arguments from opposing counsel, fallback positions if the first ask is rejected, and the strongest justification for each fallback. The model produces, in two minutes, a one-page playbook covering possibilities you would have surfaced over an hour of thinking.

The shape of the prompt matters. Vague — “how should we negotiate the cap?” — gets vague output. Specific gets specific:

We are representing a Nepali SaaS vendor. The customer (a large Indian bank) has proposed a liability cap of “12 months’ fees paid in the preceding period,” with no carve-outs. Our preferred position is uncapped liability for breach of confidentiality, IP indemnity, gross negligence, and wilful misconduct, with the 12-month cap applying only to other claims. Produce: (1) three substantive arguments we can make to the customer’s counsel for our preferred position, ranked by strength; (2) the three most likely counter-arguments we will hear, with a one-sentence rebuttal for each; (3) two fallback positions if our preferred position is rejected, with the commercial trade-off of each clearly stated; (4) one position we should refuse to accept under any circumstances, and why. Frame this for a Nepali law contract; treat customer arguments based on US/UK case law as available but not controlling.

What comes back is not the answer. It is the menu. The lawyer chooses from it — and chooses, crucially, based on what they know about the deal, the parties, and the relationship, none of which the model has access to.

Redline drafting — fact-first, then language

Once you have decided what to push for, the second high-value use is producing the redline itself. The pattern that works is fact-first: tell the model the current clause, tell it what you want the clause to do instead, ask for the redline and a one-sentence rationale you can paste into the email to opposing counsel.

Here is the limitation-of-liability clause as currently drafted: [paste clause]. We want it to do the following instead: (i) preserve the 12-month cap for ordinary claims; (ii) carve out from the cap any liability for breach of confidentiality, infringement of third-party IP, gross negligence, or wilful misconduct; (iii) make the carve-outs explicit rather than relying on a general “to the maximum extent permitted by law” formulation. Produce: (1) a clean redline of the clause achieving this; (2) a one-sentence rationale we can include in our markup or email to opposing counsel, framed as a normal commercial position rather than an aggressive one. Use the defined terms already in the contract; do not introduce new definitions unless necessary.

The model returns drafting that is almost always usable as a starting point. The “almost” matters — read it, check the cross-references, verify the defined terms — but the time from decision to draft is now minutes rather than the half-hour it used to take to write the same clause by hand.

A small technique that pays off: when the redline involves moving language around the contract (a carve-out from clause 14.2 needs a hook in clause 1.1), ask the model to list every other clause it touched. Models are reasonable at making the change; they are sometimes silently incomplete about telling you what else they changed.

Coverage of the common-counter playbook

For each of the standard contested terms in commercial contracts, there is a finite set of arguments and counter-arguments that get used. Liability caps, indemnity scope, exclusivity, MFN, IP ownership, audit rights, termination for convenience, change-of-control consent — each has perhaps five to eight standard positions on each side, and the negotiation is usually a walk through that decision tree. The model knows the tree.

A practice that some Nepali firms are starting to build: a negotiation playbook library. For each common clause type, a saved prompt that asks the model to walk the tree for your specific deal — and produces, in addition to the arguments, a “this is what we usually concede on” line that captures the firm’s own history. The first time you build this for a clause type it takes an afternoon. After that every negotiation on that clause type starts from the playbook rather than from scratch.

The library matters more than the individual prompts. Negotiation is a repeated game; the firm that captures its own pattern of arguments and concessions becomes faster and more consistent over time, in a way that no off-the-shelf AI tool can replicate.

The relationship dimension — what the model cannot see

This is where the lawyer’s role becomes irreducible. The model will, if asked, produce a maximally adversarial argument — sharp, rhetorically effective, legally defensible. You will not use it. The reason is not that the argument is wrong; it is that the parties have a five-year relationship, the opposing counsel is a friend, the client wants this deal closed by Dashain because their CFO is leaving in November, and a too-aggressive negotiation now will cost the relationship in ways the contract itself will never reflect.

The model does not know any of this. It cannot know it, because nothing in the prompt told it, and even if you told it some of it, you would not tell it everything — the rumour about the opposing counsel’s firm losing the bank mandate, the partner’s own assessment that the client is overstating their willingness to walk away, the awareness that your client’s commercial team is on its third position in two months. These are the inputs that actually shape the negotiation, and they sit in the lawyer’s head.

So the discipline is — use the model to widen the menu, never to narrow the choice. A good negotiator uses AI to ensure they have considered every plausible argument and counter; the choice of what to deploy, and how, remains theirs. A negotiator who lets the model choose for them is, in effect, negotiating without the most important information they have.

There is a related discipline on tone. Models default to a neutral-professional register that reads as cooler and more transactional than most Nepali commercial relationships actually are. A redline email that ports the model’s tone straight into the inbox can land as more confrontational than the lawyer intended. Either soften before sending, or use the model to produce the substance and write the cover note in your own voice.

Check your understanding

Quick check

—

Your senior partner asks you to prepare for tomorrow's negotiation call on a contested liability cap. The opposing counsel is a long-standing friend of the firm and the client wants the deal closed in three weeks. Which best describes the right division of labour between you and the AI?

Ask the AI to produce the most aggressive possible arguments for your position and use them as written, since the AI has seen more contracts than you have.
Avoid the AI altogether — negotiation is a human skill and using AI is a shortcut that damages credibility.
Use the AI to generate a menu of arguments, counter-arguments, and fallback positions; then choose which to deploy based on your judgment about the relationship, the timeline, and the client — and adjust the tone of any AI-drafted text before it leaves the firm.
Ask the AI to predict what the opposing counsel will do and rehearse only against that prediction.

What comes next

Contracts are the most volume-heavy use of AI in legal practice, but they are not the only one — and they are arguably not even the highest-stakes. The next chapter turns to legal research, where the model’s tendency to hallucinate authority intersects with a Nepali legal system whose primary sources are partially digitised, often paywalled, and not in any model’s training data. The citation problem is where AI in law gets genuinely dangerous, and the next chapter is about how to use the tools without getting caught by it.