Clause precedent retrieval

When a counterparty asks for a term that is not in your paper, the only useful answer is whether you have already signed something like it. Retrieval over executed contracts, plus the negotiation playbook, is how commercial counsel and deal desk get that answer during a live markup, not after a week of emailing former reviewers.

The output is a citation. Name the agreement type, the counterparty or an anonymized file ID if policy requires it, the signature date, the clause heading, and a pointer that opens the signed PDF. If the index has never seen that ask on an executed agreement of this type, write no signed precedent. Do not generate a clause to fill the silence. Counsel still decides whether to grant, fall back, or refuse.

Contract redline AI marks the live markup against the playbook. This page is the corpus check behind whether you have ever agreed to the ask.

NDAs, master service agreements, order forms, data processing addenda, and statements of work are different instruments. A limitation of liability that was acceptable in an NDA is not MSA precedent. A subprocessor list that was fine in a DPA is not a security-schedule concession on the master.

Drafts are worse than the wrong type. A redline in a workspace, a Word file named FINAL that never received a signature block, or last year's template that legal never approved is not a position you took. If retrieval cannot tell executed from working copy, do not treat hits as precedent until the repository can.

The playbook is the second corpus, not a substitute for signatures. It holds preferred, fallback, and walk-away positions. Use it to label a retrieved clause: preferred paper, a known fallback, or a one-off marked do-not-repeat. A playbook row with no signed example behind it is policy, not precedent.

Scope the search before anyone pastes the buyer's ask into a chat box:

• Same agreement type as the live negotiation.
• Executed status only. Working copies stay out of the hit list.
• Segment and date filters the playbook already uses, such as enterprise fallbacks that are not SMB paper.
• Known-bad or do-not-repeat tags, so a concession made under pressure years ago does not come back as a recommendation.

Run retrieval as a pack, not as a paragraph of uncited language.

1. Treat the counterparty ask as a question, not as a prompt to draft. "Have we signed 30-day termination for convenience on an MSA?" is searchable. "Write a 30-day termination clause" is how generated paper enters the redline.
2. Search executed agreements of that type, then the playbook positions that map to the same heading.
3. Return each hit with a file ID that opens the signed PDF, the signature date, the agreement type, the section number or clause heading, and the playbook label (preferred, fallback, or do-not-repeat).
4. If there are no hits, return no signed precedent on this ask. If the playbook has a fallback, show that row as policy, still without a fake citation.
5. Counsel, or deal desk acting inside a written grant, decides. The model does not send the markup.

Signed-contract data extraction is how party, term, and dates get out of PDFs and into fields the search can filter on. If extraction never ran, you are searching filenames.

Contract lifecycle and deal systems such as Ironclad, Sirion, Icertis, and Salesforce are a class of places signed paper and opportunity context already live. They do not create a concession you never signed. Salesforce is typically where account segment and opportunity context sit, so a search can be limited to like-for-like deals rather than the entire vault.

If policy restricts using another customer's terms as a template, show an anonymized file ID and let counsel with access open the PDF. Do not paste another customer's commercial schedule into the live redline.

Spot-check before the pack reaches the AE. Open the cited file. Confirm it is executed, the right instrument, and the quoted sentence is actually in that section. One draft cited as signed, and the queue stops trusting retrieval.

The walkthrough is illustrative, not a measured result.

A mid-market software MSA is in redline. Buyer counsel asks for termination for convenience on 30 days, with no wind-down. The playbook preferred position is 90 days with a wind-down period. Deal desk asks commercial counsel one question: have we signed 30-day termination for convenience on an MSA?

Search is scoped to executed MSAs. NDAs, order forms, and working copies are out.

The pack returns two executed MSAs that contain 30-day termination for convenience. Each hit names the file, the signature date, the section heading, and a playbook tag of fallback, used when the account was already a named exception. Counsel can open both PDFs.

The same search also finds a Word file in the opportunity folder titled MSA_v7_FINAL with 15-day termination language. It was never signed. It is excluded. Several NDAs in the vault have 30-day termination. They are excluded: wrong instrument.

Counsel now has the two signed files, the playbook fallback, and a clean statement that 15-day language was never executed. They can grant 30-day termination as a fallback, refuse, or offer a playbook middle path. Retrieval did not invent 15-day paper, and it did not treat NDA termination as MSA precedent.

On the same deal, buyer counsel also asks that the customer own all improvements to the product. Search of executed MSAs returns nothing on that heading. The output is no signed precedent. Do not synthesize an intellectual-property assignment. If an auto-drafted proposal already promised that ownership in a statement of work, that is a separate cleanup, not signed MSA paper.

Three failure modes show up in the same week if the index is sloppy.

Citing a draft as precedent. A redline, a "final" Word file, or a template is a working copy. Quote it as signed and you have told the AE that a position exists when it does not.

Mixing NDAs with MSAs. Confidentiality agreements, data processing addenda, and masters do not share risk allocation. Retrieval that ignores instrument type will surface a limitation of liability, a governing-law sentence, or a termination right from the wrong document class.

Generating a clause that was never signed. A model can produce a fluent limitation, indemnity, or assignment paragraph that matches the buyer's ask. That paragraph is not your paper. If you have never accepted the ask, say so. Offer the playbook fallback if one exists. Do not fill the gap with invented language.

Precedent is not permission. A term accepted under pressure years ago, now tagged do-not-repeat, must surface with that tag, not as a recommendation. Deal risk scoring may show the opportunity is slipping. A slipping forecast is not a reason to treat a draft as signed or to mint a clause you have never executed.

The same retrieval pattern on questionnaires is a different corpus. A security and RFP auto-responder answers from an approved knowledge base. Do not mix those answers into contract precedent. A questionnaire "yes" is not an MSA schedule you signed.

Speed here is ending the scavenger hunt: the signed file, or an honest empty result, in time for counsel to decide before the redline goes back.