A claim without a pinpoint is a rumor
Contract review is where AI assistance earns its keep fastest — and it's the use the evidence says practitioners actually run. Thomson Reuters' and Clio's 2025 survey narratives both put document review and contract summarization among the dominant real uses of generative AI in legal work, and per Anthropic's economic data, that paralegal-class layer — review, summarization, research synthesis — is exactly where usage concentrates. A model can read a 60-page counterparty MSA and hand you a clause table in under a minute. Whether that table is review support or decoration turns on one column.
The pinpoint. "Liability is capped at 24 months of fees" is a rumor about your own document. "Liability is capped at 24 months of fees — § 9.3" is a claim you can check in the time it takes to scroll. That's the entire economics of the workflow: a pinned claim costs seconds to verify against the source; an unpinned one costs you a re-read of the document, which is the work you were trying to compress in the first place. So the rule is absolute and it is the same rule as last lesson wearing a different suit: no pinpoint, no claim. An extraction row that doesn't say where it came from gets rejected unread — not argued with, not "probably fine" — and the document gets reopened. The model summarizing from its general memory of what MSAs usually say, instead of from your MSA, produces exactly this failure, fluently.
The second half of the drill is what you compare the extraction against. Not your recollection of what's market. A playbook: for each clause type, the approved language source, the preferred position, the acceptable fallback, and the walk-away — written down once by counsel instead of re-litigated in every review. This structure predates AI by decades; what AI changes is that it can now do the comparing at speed, if the positions exist as data. Three properties matter:
- Positions are set by counsel, not assembled by a tool. The playbook is human judgment, cached. AI compares the extracted clause to it; AI never gets to invent what your fallback is.
- Every entry names its source — which approved template, which version. Language in the playbook that nobody can trace is the same disease as a citation nobody opened.
- The walk-away routes to a person, by name. Anything outside preferred and fallback is a negotiation call, and negotiation is human-core work. The playbook's job is to say whose call — "escalate to counsel" with no name is a task that dies in a queue.
Put together, the review loop looks like this: extract with pinpoints → reject anything unpinned → triage each pinned clause against the playbook → escalate the outliers to their named owner. Four steps, all checkable, which means all of it can be code. Next step builds it.