Adoption is not a technology problem and never was. It is a training, incentive and measurement problem wearing a technology problem’s clothes.
Start a conversation with the AI Adoption Concierge, already scoped to adoption & change. Choose the question closest to yours, or describe your situation directly.
The gap between purchased and used is the defining feature of legal AI so far. Survey figures differ sharply depending on how the question is framed — some report use among legal professionals rising from roughly 19% in 2023 to 79% in 2024, others put lawyers actively using AI tools at around 30% in 2026 against about 11% in 2023 — but the consistent finding across all of them is that only a small minority of firms, on the order of 8%, have adopted anything universally. Most adoption is partial, uneven and concentrated in a few enthusiasts. Closing that gap is about what people are taught, how they are measured, and whether using the tool is easier than not.
What people need to know, how to tell whether it is working, and how to move beyond the first practice group.
Teach the failure modes before the features — a lawyer who cannot spot the errors cannot supervise the output.
investigateRework rate, not logins — and a baseline captured before rollout, which is the step everyone skips.
investigateWhich group first, why the prestigious one is usually wrong, and the incentive question nobody wants to open.
investigateHow the Institute approaches adoption — programme design, not a recommendation of any product.
Usually one of three things, and they are diagnosable. The tool sits outside where people work, so every use costs a context switch they pay dozens of times a day. The training covered how to use the product but not what it gets wrong, so people do not trust the output and quietly stop. Or the firm still measures and rewards on hours, so a lawyer using the tool well is reducing the number their compensation counts. The third is the most common and the least often named, because naming it means reopening the compensation model.
Failure modes first, capability second — which is the reverse of most vendor training. People need to know that citations can be wholly invented, that a real case can be cited for a holding it does not contain, that quoted text may not exist in the source, and that none of this is signalled by any hesitancy in the output. Only then does prompting technique matter. A lawyer who has been taught what to look for can supervise output safely; one who has been taught only what the tool can do will trust it exactly when they should not.
Rework rate is the single most informative measure and the least used: what proportion of output was good enough to use without substantial redoing. Licence counts and login frequency flatter every deployment and tell you nothing. Time saved matters but misleads alone, because a tool that halves drafting and doubles review has saved nothing. And whatever is measured should be measured against a baseline captured before rollout, which is the step almost everyone skips and then cannot reconstruct.
With a practice group that has repetitive, well-bounded work, a leader who is genuinely interested, and enough volume that a change shows up in the numbers. Starting with the most prestigious group is a common instinct and frequently wrong — their work is bespoke, hard to score, and their partners have the least tolerance for a tool that is imperfect. The first group's real job is to produce evidence and a story the second group will believe, which argues for choosing the group most likely to succeed rather than the one that matters most.
Describe what you deployed and to whom. The Institute will help you work out where it stalled.