Clause Extraction and Playbook Comparison
Extraction Is a Different Job From Generation
The most reliable contract work an AI system does is extraction: locating the limitation-of-liability clause, pulling the governing-law provision, listing every defined term, identifying the notice periods. The output is a pointer into a document you already possess, which means it is checkable in seconds. This is categorically safer than generation, where the output is new text with no source to check it against. The practical implication is that you should prefer workflows that return spans of the actual document over workflows that return a paraphrase. A system that says "clause 14.2 says X" and links to clause 14.2 can be verified. A system that says "the agreement caps liability at twelve months of fees" without a pointer has to be trusted or re-derived from scratch.
- Extraction returns a pointer into your own document — verification is a click, not an investigation
- Prefer tools that return source spans over tools that return only a paraphrase or a summary table
- Paraphrase without a citation to the clause is generation wearing extraction's clothes
- Extraction quality degrades on unusual drafting, scanned text, and heavily amended documents
Playbook Comparison: Encoding What Your Firm Already Decided
A playbook is a firm's accumulated position: what an acceptable indemnity looks like, which liability caps are approved without escalation, what the fallback positions are on a given clause. Comparing an incoming draft against that playbook is a well-suited task, because the standard is explicit and written down. The AI locates the relevant clause, characterises it, and flags divergence from the stated position. The value is speed on high-volume, moderate-stakes agreements — NDAs, standard vendor terms, routine procurement. The failure mode is subtle: a clause can match the playbook wording and still be unacceptable because of how it interacts with a definition elsewhere, and cross-clause interaction is exactly what clause-level comparison misses.
- Works because the standard is explicit and pre-agreed — the model is matching, not deciding
- Highest value on high-volume, moderate-stakes agreements where the playbook already covers the ground
- Blind spot: clause-level matching misses interactions with definitions and cross-references elsewhere
- A playbook that has never been written down cannot be automated — the writing is most of the work
What the Playbook Cannot Encode
Playbooks capture the standard case, and the reason lawyers are involved is the non-standard case. A clause may be within playbook tolerance and still be wrong for this counterparty, this commercial relationship, or this risk appetite on this deal. Conversely a divergence may be entirely acceptable because of something traded elsewhere in the negotiation. The model sees the document; it does not see the deal. Treat playbook comparison output as a structured list of things to look at, ranked by how far they sit from the norm, and keep the judgement about whether a divergence matters firmly with the person who knows the commercial context. Where a matter is bespoke or high-value, the playbook pass is a starting inventory rather than a review.
- A playbook encodes the standard case; the exceptions are why a lawyer is reading it at all
- Acceptability depends on the deal, the counterparty, and what was traded — none of which is in the document
- Read the output as a ranked inventory of divergences, not as a pass-or-fail verdict
- On bespoke or high-value matters, treat the automated pass as the first ten minutes, not the review
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.