Redlining a Counterparty's Contract With AI: A Playbook Approach

A small law firm can redline against its own playbook in four steps: write each clause position as a checklist (preferred wording, acceptable fallback, walk-away point), give that checklist to an AI assistant, ask it to compare the counterparty's draft clause by clause, then have a lawyer accept, amend or reject every proposed change before anything leaves the firm.

The playbook is what makes the output usable. Without it, an assistant produces generic "balanced" edits that reflect nobody's instructions. With it, the assistant is doing the part of junior review that is mechanical: finding the clause, comparing it to a stated position and drafting the nearest compliant wording. Judgement, client context and sign-off stay with you.

Follow me on Instagram@sagnikteaches

The method suits a firm of two to fifteen people, where one partner can own the playbook and one associate runs the first pass. It is general information, not legal advice, and it assumes you will apply your own professional-conduct and confidentiality rules before any client document goes into a tool. For the broader question of what AI catches and misses in contracts, read AI contract review for small firms first.

Connect on LinkedInSagnik Bhattacharya

Why a playbook beats "review this contract"

Ask an assistant to "review this supplier agreement and suggest changes" and you get a competent-sounding essay. It will flag the liability cap, mention the termination clause and suggest "more balanced" language. The trouble is that "balanced" is not a position. Your client may be a buyer who needs a cap at twelve months of fees, or a supplier who must keep it there. The same clause is a problem for one and a win for the other.

Subscribe on YouTube@sagnikteaches

A playbook removes that ambiguity. It states, for each clause type, what you want, what you will accept and what makes you stop and call the client. When the assistant has those three levels, its job changes from "have opinions" to "classify and draft". That is a narrower task, and one a reviewer can check quickly.

There is a second benefit that has nothing to do with AI. Writing the playbook forces the partners to agree on positions. Putting them on paper may show, for example, that two partners handle limitation of liability differently, and resolving that is worth the exercise before any tool is involved.

Build the playbook as a clause checklist

Keep the first version small. Pick the agreement type you see most (supplier terms, services agreements, commercial leases, NDAs) and list the ten to fifteen clauses that cause most of your negotiation. For each one, write four things in a table or a document the assistant can read.

• Preferred position: the wording or principle you open with.
• Fallback: the compromise you can accept without calling the client.
• Walk-away trigger: what must be escalated to a partner or the client.
• Drafting note: the standard replacement text you would paste in, so the assistant mirrors your house style.

Here is an illustrative extract for a firm acting for the buyer of services under a typical supplier agreement. The figures are examples, not recommendations.

Notice what is missing: long explanations. The assistant does not need your reasoning for each position, only the position itself. Reasoning belongs in a separate internal note for the people. Keep the playbook to the page or two the assistant can hold in view alongside the contract.

Set up the assistant once, then reuse it

Both ChatGPT and Claude offer Projects, a saved workspace where standing instructions and reference files stay attached to every chat. Put the playbook and the house rules for output in a Project so that each new matter starts from the same instructions. On the business plans (ChatGPT Business, Claude Team) both vendors' pricing pages say your content is not used for model training by default, and both plans start at two seats, so a sole practitioner often uses an individual plan instead. If you do, check the training setting yourself: OpenAI's help centre says "Improve the model for everyone" is on by default for personal Free, Plus and Pro accounts, and Claude's consumer plans keep the model-improvement choice in the privacy settings. Confirm the current label on the vendor's page, because settings have moved before.

Confidentiality sits above convenience. Before the first real matter, read the tool's data terms, decide what your engagement letter needs to say, and agree which documents never go in at all (for example matters under a protective order). The vendor questions are laid out in questions to ask before buying AI that touches client data.

The standing instructions in the Project can be short. This is a version you could adapt.

You are assisting a lawyer who acts for the BUYER of services.
You have the firm's playbook (attached). For every contract I paste or upload:
1. List each clause in the order it appears, with its clause number.
2. Match it to a playbook row. If no row applies, say "No playbook row".
3. Classify as: Meets preferred / Within fallback / Below fallback / Escalate.
4. Quote the exact sentence you relied on, in quotation marks.
5. For anything below fallback, draft the smallest edit that reaches the
   fallback position, using the playbook's drafting note where one exists.
Rules: never invent clause numbers or text. If a clause is missing, say
"Clause absent". If a defined term is used, say where it is defined or say
"definition not found". Do not give legal advice; flag points for the lawyer.

The "quote the exact sentence" rule is the most useful line. It lets the reviewer check any classification in seconds by searching the document for that sentence, and it makes invented text easy to catch because the quote will not be found.

Run one agreement through it: a worked example

An illustration with made-up numbers. A four-lawyer firm acts for a small logistics client buying a managed fleet-tracking service. The supplier's draft is 31 pages with 38 numbered clauses and a schedule of service levels. A junior would normally need about three hours for a first-pass markup against the firm's usual positions, and a partner another hour to review.

The associate uploads the draft to the Project. She does not ask for everything at once; she asks for the table of clause classifications first, so she can see the shape of the problem. The assistant returns a classification list. An extract of what that kind of output looks like (illustrative):

Clause 12.1 Limitation of liability - Below fallback
Quote: "The Supplier's total liability shall not exceed the fees paid in the preceding three months."
Playbook: cap at 100% of annual fees (fallback). Annual fees here: $24,000. Cap as drafted: about $6,000.
Proposed edit: replace "preceding three months" with "preceding twelve months".

Clause 14.3 Auto-renewal - Escalate
Quote: "...renews for successive 36-month periods unless notice is given not less than 14 days before expiry."
Playbook: no automatic renewal, or reminder at 90 days. Notice window under 30 days is an escalation point.

Clause 19 Governing law - Meets preferred

Two things happened that are worth noticing. First, the assistant did the arithmetic the playbook invites: a three-month cap on a $24,000 annual contract is about $6,000, which is the figure the client needs to hear. Second, it classified the renewal clause as an escalation, which means that clause goes to the partner before any markup is sent, not after.

The associate then asks for proposed edits in batches of five or six clauses, not all 38 at once. Shorter batches keep the answers precise and make review manageable. In this illustrative run, 38 clauses produced 9 below-fallback items, 3 escalations, 14 meets-preferred and 12 that matched no playbook row. Those 12 are where she spends her own time.

Total elapsed time in this example: roughly 25 minutes to classify, 40 minutes to prepare and check edits, and 45 minutes of partner review, against the four hours of human time it would usually take. Treat those numbers as a plausible shape, not a promise; your first matters will be slower while you tune the playbook.

Prompts for the clause-by-clause pass

After the classification table, work through clauses that need edits. A precise prompt gives you a drafting suggestion you can paste into the counterparty's document with Track Changes on.

Take clause 12.1 and 12.2 only. Using our playbook row for limitation of
liability, give me:
(a) the minimal tracked-style edit: show deleted words in [DELETE: ...] and
    added words in [ADD: ...] within the original sentence;
(b) one sentence explaining the change that I can put in a comment to the
    other side, in neutral language;
(c) the fallback wording if they refuse the preferred position.
Do not change any defined terms. Do not renumber.

A realistic sample of the answer, again illustrative:

12.1 The Supplier's total liability [DELETE: shall not exceed the fees paid in the preceding three months]
[ADD: in each contract year shall not exceed 150% of the fees paid or payable in that contract year].
Comment for counterparty: "A three-month cap is lower than the value of the service we are buying; a cap tied to annual fees is standard for this kind of agreement."
Fallback: 100% of the fees paid or payable in the contract year.

What you would fix: "standard for this kind of agreement" is an assertion about market practice that the assistant cannot verify. A careful lawyer would cut it or replace it with the client's own reason. Assistants like to justify edits with claims of what is "standard" or "market"; treat each one as a draft sentence to be checked, not a fact.

Minimal edits matter more than elegant ones. Counterparties respond better to a redline that changes six words than to a rewritten clause, and a short change is quicker for the partner to verify. The instruction "smallest edit that reaches the fallback" in the standing prompt does that work.

Score the risk, then decide what the partner sees first

When 38 clauses are classified, the partner does not need all of them. A short risk table at the top of the review pack directs their attention. Ask the assistant to produce it from its own classification, then check it against the clauses yourself.

Grey is the category that protects you. A playbook cannot cover every clause, and a tool that quietly passes unfamiliar text as fine is the failure that causes harm. Requiring "No playbook row" as an explicit answer forces unfamiliar clauses into the open, and a human reads those in full.

The partner review that catches what AI gets wrong

Assistants make particular kinds of mistake in contracts. The review step exists to catch them, and a short checklist is more reliable than general vigilance.

1. Search every quoted sentence. If a quote does not appear in the document, the classification beside it is unreliable. Remove it and re-run that clause.
2. Check cross-references. An edit to clause 12 can break clause 17 if 17 refers to "the cap in clause 12.1". Assistants edit in isolation unless asked to check references.
3. Check defined terms. "Fees" in the supplier's draft may exclude expenses or one-off charges; a cap on "Fees" is then lower than it looks. Ask the assistant to list every defined term inside the clause and where it is defined.
4. Read the schedules. Long attachments, service-level schedules and order forms can override the main body ("in the event of conflict, the order form prevails"). Make sure the assistant was given them.
5. Look for what is absent. Missing clauses (data return on exit, step-in rights, a termination right for repeated service failure) are invisible to a clause-by-clause comparison unless the playbook lists them.
6. Re-read the comments to the other side. Remove anything that reveals client strategy, makes a market-practice claim you cannot support, or sounds more aggressive than the partner would write.
7. Compare the final document against the original. Use Word's compare function to confirm that only intended changes are present and no text slipped in unreviewed.

Keep a record of who reviewed what. A line in the matter file such as "AI-assisted first pass against the buyer playbook v3; all edits reviewed by [initials] on [date]" costs nothing and answers the question a client or insurer may ask later.

How the same method behaves on different documents

A mutual NDA

This is the easiest case. The playbook has five rows: definition of confidential information, term and survival, permitted disclosures, return or destruction and governing law. A 6-page NDA classifies in a couple of minutes. The usual finding is a one-way obligation disguised as mutual, caught because the assistant is told to compare "who is bound" for each obligation.

A commercial lease renewal

Here the playbook matters more than the model. Rent review mechanics, repair obligations and break rights depend on the property and the client's plans, so the playbook rows say "escalate" more often than they say "accept". The assistant's value is speed in finding where the landlord's draft departs from the previous lease; give it both and ask for a clause-by-clause difference list first.

Their paper, an unfamiliar sector

Expect lots of grey rows. Use the assistant differently: ask for a plain-English summary of each unfamiliar clause and the questions you would put to the client, then decide whether the matter needs a specialist before drafting anything.

Acting for the supplier instead

If the firm sometimes acts for the supplier, keep a separate playbook rather than editing one. A single file with "buyer" and "supplier" positions side by side invites the assistant to mix them. Name the file and the Project after the side you act for.

A realistic mistake

One failure worth planning for: an assistant "fixes" an indemnity by adding the words "to the extent permitted by law", which sounds harmless and quietly narrows the indemnity. It was caught only because the reviewer used Word's compare on the final file. That is why step seven of the checklist exists.

Where the playbook comes from, and how it improves

Do not write the playbook from scratch with an assistant and trust it. Build it from your own precedents: the last five agreements of this type that you negotiated, the final positions and the emails explaining why. An assistant can help summarise those into draft rows (a good use of it), but a partner must approve each row.

After every matter, spend ten minutes updating the playbook. Record any clause the counterparty rejected and what they accepted instead; that becomes the new fallback. Record any grey clause that appeared, and decide whether it deserves a row. Over six or seven matters the grey list shrinks and the playbook starts to describe how the firm actually negotiates.

Version the file. Put "v3, updated [date]" at the top and quote the version in the matter note. If a position changes, you want to know which matters used the old one.

Firms that want the broader negotiation preparation, including fallbacks for the call itself, can pair this method with preparing fallbacks and redlines before the call. For plain-English contract reading by non-lawyers, can AI review a contract covers the basics without the playbook structure.

Three numbers to track during the pilot

Pick three numbers before you begin, so that you can tell whether the method helps and not just feels modern.

• Time to first markup. Hours from receiving the draft to a partner-ready redline, compared with your last five similar matters.
• Corrections at partner review. Count the changes the partner reverses or rewrites. If the count stays high after three matters, the playbook wording is the problem, not the model.
• Missed points found later. Anything you spot after sending (a cross-reference, a defined term). Even one is a signal to add a step to the checklist.

A reasonable pilot is three matters of one agreement type, each run both ways at least once: the usual manual first pass, then the playbook pass on the same document. Compare what each caught. Expect heavy overlap, with a few items each method found and the other missed; if that is what you see, it argues for keeping a human in the loop and for keeping the checklist honest.

When the pilot ends, decide in writing whether the method becomes standard, who owns the playbook, and what goes into the engagement letter. If you would rather map this with someone before you start, that is what my AI implementation consultation covers; you can get in touch here.

Questions lawyers ask before trusting AI with a redline

Do I need special legal AI software to redline against a playbook?

No. A general assistant such as ChatGPT or Claude can do the work if you give it the playbook in a project, paste or upload one agreement at a time, and ask for clause-by-clause output. Dedicated contract tools add Word integration and clause libraries, which matter once volume is high. Start with what you already have.

Can I put a client's contract into a consumer chatbot?

Only if your engagement terms, your professional-conduct rules and the tool's data settings allow it. ChatGPT Business and Claude Team do not train on your content by default; on personal plans, check the model-training setting yourself, because ChatGPT's is on by default. Check the vendor's terms and tell the client in your engagement letter.

How long should a playbook be for a first pilot?

Ten to fifteen clauses is enough: liability, indemnity, termination, payment, intellectual property, confidentiality, governing law, assignment and a few matter-specific ones. A short playbook that is followed beats a forty-page one that nobody reads. Add clauses after each real matter shows a gap.

Will AI produce a proper Word redline with tracked changes?

A chat assistant returns text, not a tracked-changes file. The reliable route is to have it propose replacement wording clause by clause, then type or paste each change into the counterparty's document with Track Changes on. Some dedicated tools write tracked changes directly; verify that on a test document first.

Further reads

• AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals — Clause positions for agencies that sell services to clients.
• How to Track Contract Renewals and Deadlines With AI — Catch the auto-renewal dates the redline flags.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Other routine documents a small firm can draft faster.
• Are Chat-With-PDF Tools Safe for Contracts and Client Files? — How safe PDF chat tools really are with confidential files.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Recording the time the new workflow saves.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business pricing page, OpenAI help centre (Data Controls FAQ, Projects in ChatGPT), Claude pricing page, Claude help centre (What are projects?), Claude privacy centre, and Microsoft Support (Compare and merge two versions of a document), all checked 4 November 2026. The playbook, clause table and sample outputs are illustrative.