Contract redline AI

The job is suggested markup: every counterparty edit scored against your negotiation playbook as accept, fallback, or escalate, with counter-language taken from positions counsel already wrote. Counsel still owns send. The AE does not send the file. The model does not send the file.

It is not a cycle-time claim, permission to skip liability, or an automatic outbound to the customer.

Spellbook, Sirion, and Ironclad are the class of review and CLM tools that sit next to Microsoft Word. They are not ranked here. None of them is the playbook. The playbook is the document commercial counsel maintains: preferred language, approved fallbacks, and the clause families that always escalate, including liability, indemnity, IP, data processing, most-favored treatment, audit, assignment, and governing law when it leaves your home courts.

Word is still where tracked changes live until someone files the executed copy. A model that accepts a liability cap your playbook forbids is not moving the deal. It is proposing a different contract.

Freeze the counterparty file first: filename, date, sender, and whether this is their paper or a markup of yours. Then load the playbook. For every tracked change and every inserted clause, the question is the same. Does this match a position you already take, a fallback you already allow, or something the playbook never contemplated?

"Market" is not a position. Generic training data will tell you that a low dollar cap is common on smaller orders. Your playbook will tell you whether you take twelve months of fees, a named dollar floor, and which claims sit outside the cap. Use the named floor.

Run the comparison in this order:

1. Inventory the deltas. Include tracked changes, comment bubbles that propose language, and silent inserts in a copy that arrived clean.
2. Map each delta to a playbook row: clause family, preferred, fallback, escalate-if.
3. Classify accept, fallback, or escalate.
4. Emit suggested markup in Word tracked changes, plus an issue list counsel can scan: accepted items, fallbacks with playbook cites, escalations with reasons.
5. Counsel reviews every escalate, and every accept that touches a risk family.
6. Counsel sends. Deal desk may send only under a written delegation that excludes those risk families.

The commercial paper that started this deal should already exist from auto-drafted proposal. Redline against that paper plus the playbook. Do not let the model start from a generic MSA it remembers from training.

When the playbook has no row, retrieve before you invent. Clause precedent retrieval answers whether you have ever signed that formulation. If you have not, say so on the issue list. Do not mint "standard" carve-out language because a public MSA looked similar.

Accept only when the counterparty language is inside preferred, or is more protective of you without changing the commercial bargain against you. A tighter mark on their confidentiality duties can still hide a termination-for-convenience that helps them. Read the paired clause before you classify.

Fallback when the edit is outside preferred but inside an approved alternative. Insert the fallback sentence from the playbook, not a paraphrase and not a comment that says you can be flexible on term. If the playbook's fallback is twelve months with a 90-day out after the first year, that sentence is the markup.

Escalate when the clause is on the always-escalate list, when there is no playbook row, when two clauses interact, or when this is their form and the volume of non-playbook language is high. A cap that also swallows confidentiality is an interaction. A DPA that rewrites the MSA indemnity is an interaction. Those go to counsel even if each clause looked ordinary in isolation.

Never auto-send. In a lot of email cultures, sending the Word file is an offer. Suggested tracked changes are work product. A comment that says you can live with a $50,000 cap if the deal closes this week is not executed language, and it is not an accept. The AE does not reclassify a risk-family escalate to accept in order to hit a date.

If the AE is chasing a close, the rule does not move. The issue list goes to counsel. Counsel marks send. Last-mile deal coaching can remind the rep that legal is the open step. It cannot mark the redline complete.

Overturns clustered on one clause family are a playbook gap, not a reason to raise the model's confidence threshold.

This walkthrough is illustrative, not a measured result. It claims no cycle-time saving.

An AE is working a mid-market SaaS order on the company's paper. The customer returns a Microsoft Word file with tracked changes. Payment terms and insurance limits already match preferred. Auto-renewal sits at the approved fallback. The liability section is rewritten: all claims, including confidentiality and data-protection claims, capped at $50,000.

The playbook, owned by commercial counsel, is explicit. Ordinary claims cap at twelve months of fees. Confidentiality, IP infringement, and data-breach claims sit outside that cap, or they sit above a named dollar floor that is well above $50,000. A cap that includes those claims, at that number, is a refusal.

A useful run classifies the payment and insurance edits as accept, applies the auto-renewal fallback, and puts the liability rewrite on escalate. Suggested markup restores the twelve-month ordinary cap and the carve-outs, citing the playbook rows. The issue list is short on purpose: one escalate, named, with the forbidden number in the summary so nobody has to search the tracked changes.

What almost happened: the first pass labeled the $50,000 cap as acceptable for deal size, because annual fees sat in a similar band and the model has seen plenty of small-dollar caps in other files. That is accepting a liability cap the playbook forbids. Counsel flipped the classify from accept to escalate, put the carve-outs back, and sent only after the issue list showed the refusal.

The AE did not email the Word file to keep momentum. Counsel sent.

Tracked changes are proposals. Comment bubbles are questions, arguments, or AE side-channels. The signed PDF, or the CLM fully-executed file, is the deal.

Treating a Word comment as accepted language is how a forbidden cap, a quiet assignment, or a note that you can live with unlimited audit survives into the version someone signs. If a comment proposes contract text, either promote it to a tracked change and run it through accept, fallback, or escalate, or delete it before send. Do not leave a yes sitting in the margin.

Do not send a file that still contains internal comments about walk-away, pricing, or language counsel will never take. Those comments travel with the attachment.

Clean the file on the way out: accepted changes counsel approved, rejected changes that stay rejected, comments stripped or converted. Resolving a comment is not a signature.

Playbook coverage is not complete coverage. A DPA from a regulated buyer, a most-favored-customer clause, an audit right that reaches subprocessors, assignment on change of control into an affiliate you have never reviewed, data residency you have never offered: these are unusual-risk reviews. Skipping them because the rest of the paper classified cleanly is the miss.

If the issue list is empty on a first-time industry, a first-time residency ask, or a first-time unlimited audit, the list is wrong. Route it to counsel anyway.

Deal risk scoring can flag a stalled legal cycle in the forecast. It cannot approve a clause. After signature, signed-contract data extraction is what captures the actual cap, term, and renewal so the next order form does not inherit the wrong paper from a comment thread.

Keep send in a human queue. Spellbook, Sirion, Ironclad, and Word can propose. They do not transmit.

The job is finished when counsel can open the issue list, see the playbook row for each risk clause, accept or rewrite the suggested markup, and send a file they would still defend if that cap is later in dispute.