AI is excellent at the first hour of an unfamiliar question and unreliable at the last. Build the process around that shape.
Start a conversation with the AI Adoption Concierge, already scoped to research workflow. Pick a starting point, or describe your situation directly.
The productive way to think about AI in legal research is as a very fast, very confident, occasionally wrong research assistant who has read everything and understands none of it deeply. That assistant is enormously valuable for orientation — what is the shape of this area, which lines of authority exist, what should I read — and cannot be trusted for the final answer. A workflow built on that reading uses AI at the start and the human at the end, with verification to primary sources as the non-negotiable step between. The version that fails is the one where AI produces the answer and a human skims it, because skimming is exactly the review that fluent-but-wrong output defeats.
AI is strong at the top of this list and progressively weaker down it.
The strongest use. What the landscape looks like and what to read next.
Surfacing cases and provisions worth examining. Fast, and a starting set rather than an answer.
Useful with the caveat that summaries overstate. Read the ones that matter.
Genuinely helpful for structure, and where unsupported synthesis starts to appear.
Weak. This is judgement, and it is where confident wrongness is most costly.
A citator step and human confirmation. Not a place to rely on generated output.
How the workflow is built.
Whether the speed is real, and whether the firm ends up in a reported decision.
The wrong turns and the reading around are where understanding forms. A junior handed a synthesised answer skips that — and cannot tell when the synthesis is wrong. Requiring verification to primary sources preserves most of the learning and most of the speed.
Open the authority. Not the tool's summary of the authority, not the snippet in the answer — the document, at the passage relied on. Almost every reported failure and most near misses come down to somebody accepting a characterisation without reading the source. It is a small amount of work relative to the research it replaced, and it catches both categories of error: the case that does not exist, and the real case that does not say what the answer claims.
Lightly, and consistently. A short note on the matter recording that authorities were verified, by whom and when, is enough for most purposes and is the difference between describing a process to a court and having evidence of one. Some firms attach it to the research memo template so it happens by default. What does not work is relying on people to remember to record it, because the occasions when it matters most are the occasions when everyone was busiest.
It should, and most firms have not addressed it yet. If juniors use these tools — and they will — the skill to develop deliberately is evaluating a synthesised answer against primary sources, which is a different competence from constructing research from scratch. Some firms have juniors research a question both ways early on, precisely so they experience where the tool was wrong. That is expensive in time and it produces lawyers who know what they are checking for, which is the whole point.
Exactly where it always did, and it is more necessary rather than less. Generated answers can rely on authority that has been reversed, superseded or distinguished, and unless the tool genuinely integrates subsequent history — which should be verified rather than assumed — nothing in the AI step catches it. Treat the citator check as a separate, mandatory stage rather than something the research tool has handled, unless you have specifically confirmed that it has.
Describe how research runs at your firm. The Institute will help you design around the error rate.