The E-Discovery AI Review Strategist Owns the Validation, Not the Review

What Does an E-Discovery AI Review Strategist Decide Once the Platform Ranks?

A vendor demo ends and the general counsel asks the only question that matters. If opposing counsel challenges this production, what do you hand the judge? The platform's confidence ranking is not an answer. What answers is a written protocol, a validation sample somebody actually drew, and a lawyer willing to certify the result. Deciding all three is now the seat.

Generative tools have landed hardest on exactly the tasks that used to absorb the contract review budget: first-pass responsiveness, quality control, and privilege assessment, which shortens discovery timelines and moves legal staff toward higher-value case work ¹. No public survey puts a reliable number on how far review headcount has fallen, and the percentages circulating in vendor material are not measured data. The direction is not in doubt. As the linear review room shrinks toward a supervising handful, the risk model changes shape. It no longer rests on averaging out individual reviewer error across a large room. It rests on whether the protocol was sound and whether anyone checked.

So the decisions that remain are specific and they are all judgment calls. Which prompt or classifier definition counts as the operative instruction, and does it match the request for production's actual language. What recall target the production has to reach before it ships, and on what sample size. Whether privilege gets a separate model, a separate sample, and a separate second reviewer, since a false negative there is not a rework item but a waiver argument. What gets disclosed about the methodology, and when, because a protocol negotiated up front is defended differently than one revealed after a motion.

In United States federal practice, Rule 26(g) of the Federal Rules of Civil Procedure requires an attorney to sign discovery responses, certifying after reasonable inquiry that they are complete and correct as of the time made. Nothing about a machine doing the first pass moves that signature. Which inquiry is reasonable for a given production is a case-specific question for counsel, and practice varies by jurisdiction and by judge. The point for a hiring manager is narrower: someone has to be able to describe, in ordinary sentences, what inquiry was made.

Which Tells Show That a Candidate Has Defended a Production, Not Just Run One?

Go back to the general counsel's question. A candidate who knows the software answers with the ranking; a candidate who has been on the wrong end of that question knows what the ranking would have to look like to be wrong, and has checked. Ask any of them how they would find out that a responsiveness model was quietly missing a document family, and listen for a sampling design rather than a feature name.

Six tells worth watching for:

• They ask what the request for production actually says before discussing the tool. Responsiveness is defined by the request, not by a topic the model finds interesting, and a strategist who starts at the platform has skipped the only source of truth.
• They distinguish recall from precision without prompting, and say which one they would trade. Over-inclusion costs money. Under-inclusion costs the case. A candidate who treats those as symmetric has never had to defend a production.
• They treat privilege as its own workflow. Separate model, separate sample, second-pass human review, a log of every downgrade. Folding privilege into general responsiveness QC is the single most common protocol error and a candidate who does not flinch at it is not the hire.
• They can describe an elusive document. Ask for one the model missed and why. Real answers are textured: a chain where the privileged content lived in the quoted tail, an attachment whose parent was responsive and whose child was not, a spreadsheet whose relevance sat in a formula.
• They keep a decision log by reflex. Every protocol changes mid-matter. The defensible version records what changed, when, and what was re-run.
• They will say the number is not ready. Someone who has never held up a production has never been in the position the seat exists for.

One anti-tell. A candidate who offers to detect which documents were written by AI is selling a capability that does not exist and does not belong in a discovery protocol. The honest work is documenting how the review was conducted, with named humans accountable at each step.

Which Backgrounds Produce an E-Discovery AI Review Strategist?

The direct feeder is a litigation support or discovery counsel background that already ran technology-assisted review, since predictive coding taught the same core motion a decade earlier: define the target, train, sample, measure, document, defend. Those people mostly need a semester of hard reading on how generative models fail differently. The useful part is that they already argue about recall for a living.

The less obvious backgrounds are the ones worth interviewing. Statisticians and survey methodologists bring sampling instincts that most lawyers approximate; a person who can size a sample for a stated confidence interval and explain it to a judge in two sentences is rare and disproportionately valuable. Regulated-industry quality assurance, particularly pharmaceutical or device QA, produces people who treat a change to a validated process as an event requiring re-validation rather than a routine update. Forensic examiners bring chain-of-custody habits and a healthy assumption that the collection itself is where the real gaps hide. Career contract attorneys who spent years on privilege queues bring the pattern library nobody else has, since they have read the ten thousand borderline documents that make an abstract protocol concrete.

How the strong ones got good is visible if you ask. The answers that mean something are about checking rather than using: running the same privilege question through a model three ways to see where the calls diverge, asking a model to summarize a protocol and then finding the two obligations it invented, keeping a running file of every instance where a confident classification turned out wrong and what the tell had been. That is the same discipline that shows up in adjacent oversight seats, where the person's value is catching what the automated pass missed rather than matching its throughput, much like a complex case escalation specialist working the queue that the automation hands back.

Candidates who avoid these tools entirely misjudge which parts are hard. Candidates who trust the output fail worse, because a fabricated cite in a meet-and-confer letter is a credibility loss that no amount of later accuracy repairs.

Source Review Strategists Where Validation Protocols Already Get Argued

Post where people already defend a methodology in front of someone hostile. The e-discovery bar has real institutions: EDRM, the Sedona Conference working groups, ACEDS and its certification community, and the regional e-discovery meetups that cluster around federal court districts. Those rooms are full of people who have argued about recall targets in a hallway. A general legal job board returns people who have used the software.

Feeder employers follow the same logic. E-discovery service providers and managed review companies are moving senior people into protocol and QC roles as the review floor shrinks. Large firms' in-house litigation support groups have carried this function for years under an older title. Regulated corporations with steady investigation volume, meaning banks, insurers, pharmaceutical companies and large healthcare systems, run internal discovery teams whose members leave for scope rather than for pay. Government enforcement and inspector general offices produce people used to defending a process on the record.

Search on the duty, since the title has not settled. Litigation support manager, discovery counsel, review manager, TAR project manager, e-discovery consultant and legal operations manager all show up covering some part of this work. The brief you write matters more than the noun you pick.

Screen on artifacts before you screen on interviews. Ask for a review protocol, a validation memo, a privilege QC plan or a meet-and-confer position paper, redacted as far as needed. This is a writing job that happens to involve software, and the writing sample settles in fifteen minutes what three interviews will circle. Read it for whether the reasoning survives an adversarial reader, not for polish.

Authority to Hold a Production Closes This Offer More Often Than Pay Does

Close on standing. The candidates worth hiring have watched a production ship over their objection because a partner needed the date, and they read your offer against the general counsel's question: can the person who answers it also stop it. Name who they report to, what they can hold, and how an override gets written down. A seat responsible for defensibility with no power to stop a production will be blamed for an outcome it could not change.

On pay, be honest about what is knowable. As of September 2026 no government wage series covers this title, and the salary numbers circulating in vendor content are not survey data. What is observable is directional: what these seats are bought for is judgment and defensibility rather than review volume, which pushes these roles toward the top of the litigation support band rather than the middle of it, and firms deploying discovery AI are explicitly working out how to do so defensibly with attorney oversight intact ². Benchmark against your own senior litigation support and discovery counsel postings and against what the regional service providers are paying for protocol and QC leads. Do not quote a national figure you cannot source; the candidates in this market will know you made it up.

What kills the offer is predictable. A description that turns out to mean running the platform. A reporting line through the team whose deadlines the protocol constrains. No budget for the validation sample, which reads immediately as a statement about how seriously the mandate is meant. And a hiring process slower than the vendor's roadmap, since a candidate watching the tooling change every quarter is unwilling to spend four months on a process.

On location, most of this travels. Protocol drafting, sampling design, QC review and meet-and-confer preparation are all remote-native, and the review team itself has been distributed for years. Three things do not travel. Client data under a protective order or a restricted-access agreement often has to be reviewed inside a controlled environment, which can mean a secure room and named personnel. Depositions, hearings and some regulator interactions are in person. And the first month is discovery about your own discovery, meaning sitting with IT and custodians to learn where data actually lives, which goes badly over video with people who have no reason to trust you yet. The same pattern shows up wherever oversight of an automated process is partly remote and partly physically anchored, as with an AI voice support operations specialist who can audit from anywhere but has to sit with the floor to learn what the transcripts leave out. Write remote with scheduled on-site weeks into the offer rather than negotiating it later.

Read the evidence >