Roles
What Is An Applied Legal Researcher, And Who Should You Hire?
An applied legal researcher is a practicing-level lawyer employed by a legal AI company to define what correct output looks like and to judge whether the model produced it. The work is neither practicing law nor writing code. It sits inside the model team, building evaluation sets, answer keys and failure taxonomies. Harvey lists the title on an otherwise engineering-heavy board [1], and the category is still forming, so the title moves between companies.
The takeMost legal AI teams discover this role backwards. They ship, a customer's partner finds a wrong citation, and someone realizes nobody on the build side could have said in advance what right looked like on that document. Review capacity does not fix that, because a reviewer catches one output at a time while the answer key catches a class of them. The hire that matters is the lawyer willing to stop practicing and start writing down the judgment they used to apply case by case. That person is rarer than a good reviewer and worth more to the model than three of them.
Where Olive fits
Open a role and see what the work shows
Olive is priced per attempt rather than per seat, and an attempt returns six evidenced findings on one candidate: an input to your decision, never a ranking or a filter. Ten attempts a month are free, so a pilot can run beside your current round and be compared against it.
Rank your shortlistWho Writes The Answer Key When The Model Drafts A Motion?
A model returns a summary judgment brief that reads well, cites six cases, and gets the standard of review subtly wrong. An engineer cannot see it. A reviewing attorney catches it on that one document and moves on. Nobody has written down what the model should have done, so the next release regresses in the same place and no test fails. That gap is the job.
An applied legal researcher turns professional judgment into something a build process can consume. In practice that means a graded set of real tasks with authored answers, a written rubric for what separates a passing draft from a failing one, and a taxonomy of failure modes specific enough to act on: fabricated authority, real authority mis-stated, correct law applied to the wrong procedural posture, a hedge upgraded into a conclusion. Each of those needs a different fix, and only a lawyer can tell them apart at speed.
The deliverables are unusual for a lawyer. An evaluation set of a few hundred graded examples across a practice area. Annotation guidelines other lawyers can follow without calling you. A weekly ruling on the disagreements those annotators escalate. A short memo to the research team saying which of last sprint's improvements was real and which moved a number without moving quality. Some of them sit with engineers reading model outputs for an afternoon, which is closer to the work of an AI engineer than to anything in a practice group.
Which Traits Separate A Real Applied Legal Researcher From A Lawyer Who Likes AI?
The separating trait is specificity about wrongness. Hand a candidate a model-drafted memo with a plausible error in it and ask what went wrong. A strong candidate names the error, names the class it belongs to, and says how they would test for that class. A weaker one says the model needs attorney review, which is true, general, and worth nothing to a build team.
Four more tells hold up. They will commit to an answer key under disagreement, saying which of two defensible drafts is better and writing the reason down, because a rubric that dissolves under argument produces no gradeable data. They distinguish a legal error from a stylistic preference without being coached, and they say out loud when a task has no single right answer rather than inventing one. They think in samples rather than anecdotes, asking how often before asking how bad. And they will tell you which of your model's current outputs are already good enough to ship, which is the tell that they will not simply mark everything unsafe and call it rigor.
Performed fluency has a signature too. Vocabulary is current and unattached to a graded example: hallucination, retrieval, agents, all correct, none of it connected to a document they marked wrong for a stated reason. The candidate proposes a governance framework when asked for a rubric. And they treat the role as a step toward general counsel work, which is a different job with a different daily rhythm, closer to what an AI product counsel does in design review.
Which Backgrounds Produce This Person, And How Did They Get Good With AI?
Three sources, and the third one surprises hiring managers. Mid-level associates from litigation or transactional practice, four to eight years in, who already spent their days deciding whether a junior's draft was right. Lawyers who moved into knowledge-management work, where writing standards for other lawyers is the whole function. And people from legal publishing and headnote editorial work, who have written classification standards at scale.
Other paths land too. Clerks who read for a judge and learned to say why one brief beat another. Bar-exam and casebook writers, who have already built graded answer keys for a living. Compliance and regulatory specialists whose instinct for evidence trails transfers directly, in the way it does for a clinical AI deployment and model drift auditor in medicine.
How the good ones got good is consistent, and it is not a course. They used an assistant on their own matters until they had a mental catalogue of where it fails them. They started asking for the source before reading the sentence, because they got burned once. They noticed that the model is strongest on structure and weakest on posture, jurisdiction and dates, and they can tell you that without prompting. Ask what they stopped delegating after a bad experience. The answer should be specific to their practice, dated, and slightly embarrassing. A candidate whose AI use has produced no scar tissue has been reading about it rather than working with it.
Where Do You Find Applied Legal Researchers, And How Do You Close One?
Look where lawyers have already left practice for a build. The visible employers today are the legal AI vendors themselves, Harvey among them 1, plus the evaluation teams inside the large legal publishers. Their alumni are the warmest pool you have. Beyond that, knowledge-management teams at large firms, legal-tech groups at law schools, and bar-association technology committees hold people already doing part of this job under an older title.
Search by work rather than title, because the category is still forming and the same role appears as applied research, legal domain expert, legal knowledge engineer, or subject-matter expert on a model team. A posting that says applied legal researcher and one that says legal quality lead can describe identical work, and candidates screen themselves out on the word.
What closes them is intellectual authority, stated concretely. This candidate is leaving a partnership track, and they want to know their ruling on a disputed answer stands. Say who overrules whom. Name the research lead they will sit with and the cadence at which their rubric reaches the training pipeline. Show them a real disagreement your team had last month and how it was settled. What kills the offer is the shape of a review queue: a job description that promises research and delivers a daily quota of outputs to check reads as document review with better branding, and the strongest candidates leave the process for exactly that reason.
What Does This Role Pay, And Does It Sit In The Office?
No published salary series covers this title, so treat any precise number for an applied legal researcher as inference rather than data. The honest framing is which band you are hiring against, and it is the mid-to-senior in-house attorney band rather than any research or annotation band. These candidates are recruited out of practice, and the offer has to survive comparison with the compensation they are leaving. Teams that price the role like contract review do not fill it.
Two forces push the band up. Equity does real work, because the person is accepting a title that is not yet a recognized rung on a legal career ladder, and upside is how that risk gets paid. And AI-skilled roles carry a measurable premium: PwC's 2026 AI Jobs Barometer, drawing on close to a billion job advertisements, reports an average wage premium of 62 percent for jobs demanding AI skills 2. That is an average across occupations rather than a figure for this title, so use it as a direction, not a number. Anything more precise should come from your own comparables, gathered this quarter.
On location, the work is largely remote-capable and one part is not. Writing an answer key is solitary. Settling annotator disagreements and calibrating with the research team benefits from a room, and teams that run this fully asynchronously tend to discover drift in their own rubric a month late. A workable norm is remote with a scheduled calibration session and periodic on-site weeks. Client-confidential corpora impose their own answer: some work has to happen on managed machines in a controlled environment, and saying that plainly in the posting saves everyone a late-stage surprise.
Common questions
How do I become an applied legal researcher?
Build evidence that you can convert judgment into a standard other people can apply. Inside practice, volunteer for knowledge-management work: write the precedent guidance, the drafting checklist, the review rubric your team argues over. Use an assistant on real matters until you can name where it fails in your practice area, with examples. Then build a small public artifact: fifty tasks in your specialty with authored answers and a written rubric, plus a short note on what the model got wrong and how the errors cluster. That artifact does more in a hiring conversation than a resume line, because it is the actual deliverable of the job.
How is this different from having lawyers review model output?
A reviewer catches one wrong output. An applied legal researcher catches a class of them by writing the standard that a test can run against. Review is per-document and does not accumulate, so the same failure returns in the next release with nothing to flag it. The researcher's output is durable: graded example sets, annotation guidelines, a failure taxonomy the engineering team can act on. Most teams need both, but they are separate hires with separate skills, and a job description that blends them usually ends up as a review queue with a research title on it.
Does this person need a bar license and practice experience?
Practice experience is close to essential; an active license is often not, and requirements vary by employer and jurisdiction. The role is not the practice of law, so it usually does not require licensure the way advising a client does, but the judgment being encoded comes from having done the work under real stakes. Some employers still ask for an active license for credibility with customers or for adjacent duties. Because the line between this work and legal advice is jurisdiction-specific and moves, confirm the scoping for your own posting with counsel rather than copying another company's language.
What should the take-home or interview exercise be?
Give them a real model output from your product with a genuine defect and forty minutes. Ask three things: what is wrong, what class of error is it, and how would you write a test that catches the class next time. Then ask for a rubric for one narrow task and see whether it survives a disagreement you introduce. Avoid the pure writing sample. A polished memo is exactly the artifact a model produces well, so reading one tells you little about the person who submitted it. Watching the judgment happen is the part that separates candidates.
Is the title stable enough to hire against?
Not yet. The same work appears as applied legal researcher, legal domain expert, legal knowledge engineer, and subject-matter expert on a model team, and the visible postings are concentrated among legal AI vendors and legal publishers rather than spread across the market 1. Two practical consequences. Write the posting around deliverables, an evaluation set, a rubric, a failure taxonomy, so candidates recognize the work whatever they call it. And expect to source by outreach rather than by inbound, because the people who would be excellent at this are not searching for a title that has not settled.
How many of these should a team have?
One per practice area the product claims to serve is the pattern that holds up, not one per company. Litigation, transactional, and regulatory work fail in different ways, and a rubric written by a transactional lawyer will not catch a procedural-posture error in a brief. Small teams start with one person covering the area that carries the most customer risk, then add annotators the researcher trains and adjudicates for. The ratio that tends to work is one researcher setting the standard for several reviewers applying it, rather than a flat pool of lawyers each judging by their own instinct.
References
- 1. Harvey job board (Ashby posting API) ✓ api.ashbyhq.com Discovery sweep found the title Applied Legal Researcher listed on an otherwise engineering-heavy board, placing a practicing-level lawyer inside the model team.
- 2. PwC 2026 AI Jobs Barometer pwc.com Reports an average 62 percent wage premium for jobs demanding AI skills across close to one billion job advertisements. An average across occupations, not a figure for this title.
2 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.