Some tasks compress by most of their duration. Others barely move, and a few get slower once verification is counted. Knowing which is which is the whole exercise.
Start a conversation with the AI Adoption Concierge, already scoped to productivity & workflows. Choose the question closest to yours, or describe your situation directly.
Productivity claims in legal AI are usually stated as a single number, and a single number is the wrong shape for the answer. What actually happens is that a matter is made of many tasks, some of which compress dramatically and some of which do not compress at all — and the ones that do not tend to be the ones that determine how long the matter takes. First drafts get much faster. Reading and thinking about what the draft should say does not. Finding candidate authority gets much faster; confirming it says what you think does not. Firms that measure at the task level get useful, actionable numbers and can redesign the work around them. Firms that measure at the matter level get a disappointing number and conclude the technology does not work, which is the wrong conclusion drawn from real data.
The three places where the time actually is, and what happens to each.
Where drafting assistance genuinely helps, and the failure mode that looks like success.
investigateThe clearest productivity case on the site — and the one with the sharpest sampling question.
investigateDiscovery, depositions and case preparation — where the court is watching and the standing orders apply.
investigateHow the Institute approaches productivity — task-level, measured, and honest about what did not move.
Large on a narrow set of tasks and modest across a whole matter, which is why the headline figures and firms' lived experience diverge so often. First-draft production, document summarisation, initial research orientation and bulk review can compress by most of their duration. Reading, judgement, client communication, negotiation and verification barely move. Since a typical matter is a mix, the blended saving is usually well below the task-level maximum — and still worth having. The firms that are happiest with their results are the ones that expected this shape in advance.
High-volume, low-variance, well-specified tasks — the work that was tedious rather than difficult. Bulk document review, summarising large record sets, first drafts of routine documents, extracting structured information from many similar files. The common feature is that a competent person could do the task but it takes them a long time and the output is checkable. Where the task requires deciding what should happen rather than producing something, gains are much smaller regardless of the tool.
Yes, in two identifiable situations, and both are avoidable once named. The first is a task that needed judgement more than production, where the lawyer now reads and corrects a plausible wrong answer instead of writing the right one — correcting a bad draft can take longer than drafting. The second is verification overhead exceeding the production saving, which happens on short tasks where checking costs nearly as much as doing. Both argue for choosing tasks deliberately rather than applying AI uniformly and hoping.
At the task level, on the full task, against a real baseline. Full task means to a usable, verified result — time-to-first-output is the number vendors quote and it is not the number that determines whether a matter finishes sooner. A real baseline means comparable prior work, ideally recorded before adoption, because retrospective estimates of how long things used to take are unreliable in a consistent and flattering direction. Small and rigorous beats large and impressionistic here.
Describe a matter type. The Institute will help you find the tasks that compress.