What partners reject: a taxonomy of declined redlines
We coded 184,000 rejected redlines where the reviewer gave a reason. A third were correct and not worth the fight. Fewer than one in five reflected a misreading of the playbook. The rest are more interesting.
When Review 4.0 shipped in July 2025, it added a small thing: a rejection reason. A reviewer declining a redline could pick from a short list or type a line of her own. We added it so the system could learn within the matter. It has had a second use we did not plan for: it tells us, at scale, why lawyers say no.
This note reports what we found in the reasons recorded between January and June 2026 across active accounts, anonymised and aggregated: a taxonomy, with proportions, method and the caveats that optional free text deserves.
Sample and method
Over the six months, reviewers rejected roughly 410,000 proposed redlines. Of those, 184,000 carried a reason, either a selection from the list or free text or both. The remaining 226,000 were rejected silently, and we say more about them below.
We drew a stratified random sample of 6,000 reasoned rejections, weighted by account so that no single large firm supplied more than 3 per cent of the sample. Two lawyers on our team independently coded each rejection into one of six categories, reading the reason, the redline, the clause and the playbook rule cited. Where they disagreed, on 11 per cent of items, a third lawyer resolved it. Agreement before resolution, as Cohen’s kappa, was 0.79, adequate for six categories with fuzzy boundaries.
The proportions below are from the coded sample applied to the full 184,000, with sampling error of roughly plus or minus one point on the larger categories.
The taxonomy
Correct, but not worth the fight: 34 per cent. The reviewer agreed that the clause departed from the playbook and that the redline would fix it, and chose not to raise it. The reasons were commercial: the counterparty would not accept it, the point was minor relative to the deal, the relationship mattered more. This is the largest category, and it is not a failure of the system. It is a judgement the system is not asked to make.
Right issue, wrong fix: 21 per cent. The flag was correct; the proposed language was not what the reviewer wanted. Sometimes the redline was too aggressive for the stage of negotiation, sometimes it used a formulation the firm did not favour, and sometimes it fixed the clause in isolation when the reviewer intended to address it elsewhere. This is the category most directly about redline quality, and the one where in-matter learning has the most effect.
Playbook misread: 18 per cent. The system applied a playbook rule that did not apply, or applied it wrongly. The most common pattern was a rule with an implicit condition, usually deal size or counterparty type, that the playbook did not state. The second most common was a rule from the wrong practice area’s playbook being applied because the document was misclassified. Fewer than one in five rejections falls here: lower than we expected, higher than we would like.
Already addressed elsewhere: 12 per cent. The flagged clause did depart from the playbook, but another clause in the document, or a side letter, or a schedule, cured the departure. Review reads the whole document, but reads side letters only when attached to the matter, and is less confident about cross-references between schedules and the main body.
House style and tone: 9 per cent. The redline was substantively acceptable and the reviewer rejected it to rewrite it in the firm’s or the partner’s preferred voice. Draft’s voice modelling addresses this for composed documents; Review’s redlines have had less of it, and this category is the reason we have been testing voice-aware redlines in Review since the spring.
Other: 6 per cent. Reasons we could not classify, including “no”, “discussed with client” and a handful that appeared to be typed into the wrong field.
What the silent rejections hide
The 226,000 rejections without a reason are not a random subset. Accounts with high rates of silent rejection also have shorter time-on-document and lower rates of accepted-after-edit, which suggests reviewers working fast and not engaging with the proposal. We suspect silent rejections are disproportionately in the first category, correct but not worth the fight, because that is the judgement a reviewer makes fastest and feels least need to explain. But we cannot show it, and the taxonomy should be read as describing reasoned rejections only.
What changed as a result
Three things, so far.
The playbook editor now prompts for a condition whenever a rule is entered as absolute. “Always” and “never” trigger a question: does this apply regardless of deal size, counterparty type and document type? Early cohorts since the change show fewer playbook-misread rejections in their first month, though the sample is small.
Review 4.2 reads attached side letters and schedules in the same pass, and the flag now shows which other clause it considered and why it did not treat it as curing the departure. Reviewers can correct that judgement, and the correction is learnt.
And we added “not worth raising” as an explicit option in the rejection list, separate from “disagree”. The two had been conflated, which made the system treat a commercial judgement as evidence that the playbook was wrong.
Caveats
Rejection reasons are optional and self-reported, and reviewers under time pressure pick the first plausible option. The coding was done by our own lawyers, who know the system and may read ambiguity in its favour. The six-month window begins after Review 4.1 and includes the 4.2 release in March, so the system under study changed partway through. And the proportions are across all document types; an NDA-heavy account will show a different mix from one reviewing SPAs.
The finding we keep returning to is the first one. A third of the time a partner says no to a correct redline, she is right to, and the system’s job in that moment is to have made the point clearly enough that saying no was an informed decision.