Roles
A Learning Data Analyst Earns The Job By Killing Your Favorite Course
A Learning Data Analyst turns AI tutor logs, LMS events and assessment records into decisions about what to keep teaching and what to stop. The work is joining learning data to work outcomes, maintaining the skills taxonomy the rest of the organization plans against, and checking AI-written analysis before a leader repeats it. Hire the person who can tell you a well-liked program changed nothing, and show the join that proves it.
The takeMost learning analytics hires fail because the role is scoped as reporting. Dashboards get built, completion climbs, and nobody can say whether anyone works differently. The job only pays for itself when the analyst is allowed to reach outside the learning platform, into ticket resolution times, error rates, ramp curves, quality scores, and to publish a finding that embarrasses a program someone senior sponsored. If the analyst cannot get that join, or cannot survive publishing it, the seat is a reporting seat with a better title. Hire for the reach, then actually grant it.
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 shortlistYour LMS Reports Ninety-Four Percent Completion. Now What?
Ninety-four percent completion, four-point-six average satisfaction, eleven thousand AI tutor sessions last quarter. The CFO asks what changed on the floor and the honest answer is that nobody knows, because nothing in the learning stack is joined to anything people actually do at work. That is the moment this role exists for.
A Learning Data Analyst starts by refusing the platform's own definition of success. Completion is attendance. Satisfaction is a mood taken at the end of a session, often by the tool being evaluated. The analyst's first project is usually unglamorous plumbing: get learner records keyed to the same identifier as the operational systems, agree what counts as an event, and establish a baseline for the behavior the training was supposed to move.
Then the interesting part. AI tutors changed the raw material of this job. A tutor session is a transcript, not a score, and it records where a learner got stuck, what they asked twice, and where the model gave an answer that let them stop thinking. That is far richer than a quiz result and far harder to aggregate, which is why the role is opening now and why generic BI analysts struggle in it. Reading a thousand tutor conversations for recurring confusion is closer to qualitative coding than to SQL, and the person you want is comfortable doing both in the same week.
Screen the Learning Data Analyst on What Changed, Not What Was Measured
Ask for one program the candidate measured and one decision that followed. The tell is whether the story ends in a stop, a rebuild, or a redirected budget. Weak candidates narrate methodology and dashboard architecture. Strong ones narrate a fight: the join that was missing, the proxy they settled for, who disagreed with the finding, and what happened next.
Three specific probes separate real from performed. First, ask how they would know a training program did nothing. Most people describe a positive result and how they would detect it; the analyst you want describes a null result and how they would defend one, including the confounds that make a null hard to publish, such as seasonal volume, a tooling change shipped the same month, or the fact that keen learners were already the better performers before they enrolled. Selection effects are the default failure mode in this data and a candidate who does not raise them unprompted has not done the work.
Second, hand them your actual skills taxonomy, or the spreadsheet standing in for one, and ask what breaks. Look for someone who immediately asks how many levels it has, who is allowed to add a skill, and how last year's definitions compare to this year's, because a taxonomy that gets quietly redefined every quarter cannot support a trend line and the whole skills-based planning effort rests on it. About forty percent of core job skills are expected to change by 2030 and sixty-three percent of employers name the skills gap as their biggest barrier to transformation 1, which means the taxonomy is not documentation. It is the instrument.
Third, ask what they would refuse to report. A person who has never withheld a number has never been leaned on. The answer worth hearing names something concrete: no individual-level learning scores to line managers, no per-person leaderboards, no engagement metric presented as competence. Learning data is behavioral data about employees, and an analyst without instincts there will build something you have to dismantle later.
Which Backgrounds Actually Produce This Person?
The obvious pipeline is people analytics, and it half works. People analytics folks bring the HRIS joins, the privacy instincts and the executive audience, but many have never opened a learning transcript and default to survey instruments. The second obvious pipeline, L&D itself, produces the opposite gap: real instructional judgment, thin data engineering.
The unexpected sources are better than either. Institutional research staff from universities have spent years matching intervention data to outcome data with weak identifiers and strong scrutiny, which is precisely this problem. Clinical trial data managers and program evaluators in public health arrive with the null-result discipline you cannot teach quickly. Assessment psychometricians know exactly why a fifteen-item quiz does not measure a skill. Contact center quality analysts have already read thousands of transcripts for recurring failure patterns, which transfers directly to AI tutor logs and overlaps with the work of a learning content quality reviewer.
What none of those backgrounds guarantees is fluency with the AI layer, and that is now table stakes rather than a bonus. Data and AI capability is being named among the essential skills of the L&D function for 2026 3, and the roles expanding fastest in learning and edtech are the data-facing and AI-enabled ones 2. If you inherit an analyst from a reporting team, budget the first quarter for learning the domain, not the tools.
Watch How They Check an Analysis a Model Wrote
This candidate uses AI heavily and should say so without hedging. The distinguishing question is not whether they use it but what they do in the ninety seconds after it produces a confident paragraph. Ask them to walk through a specific instance where a model's output was wrong in a way that would have survived a casual read.
The good version of this person has a routine. They ask the model for the analysis and then reproduce the number by hand on a subset before it leaves their machine. They watch for the failure that matters most here: a language model asked to interpret learning data will narrate a plausible causal story from a correlation, because that is the shape of the writing it has seen, and the story arrives fully formed with the caveats already smoothed off. An analyst who pastes that into a leadership deck has automated the exact error the role was created to prevent.
They also use AI where it genuinely earns its place. Clustering thousands of tutor conversations by the concept a learner struggled with is real work a model does well, as is a first pass at mapping free-text job postings onto taxonomy skills. Both need a human sample audit before anyone acts, and the candidate should describe the audit as part of the method rather than as a nice-to-have. The same verification instinct is what separates a competent AI tutoring oversight coordinator from someone who just watches dashboards, and the two roles end up reviewing each other's evidence.
In interviews, do not ask about AI use in the abstract. Give them a short analysis a model wrote, with one real flaw and one cosmetic imperfection, and ask what they would send to a VP. The candidate who fixes the typo and forwards the causal claim has told you everything.
Close Them With Access and Honest Comp Framing
What closes this hire is reach, not title. The offer that wins names the systems they will be granted, the operational data they can join against, and the person senior enough to defend a finding nobody wanted. Candidates who have been burned before will ask directly whether a negative result has ever been published internally, and a vague answer loses them.
On compensation, no reliable public benchmark for this specific title was found for this piece, so treat any single figure you see quoted as unsourced. What is defensible is the relative anchor: the role competes with people analytics and product analytics for the same profile, and pricing it against L&D coordinator bands is the most common way these searches stall for months. Get a current, dated market pull from your compensation team before posting, and calibrate against the analytics bands rather than the training bands.
The offer usually dies in one of three ways. The role reports into a function that cannot see operational data, so the work is impossible on day one. The scope turns out to be dashboard maintenance for an existing tool. Or the taxonomy is owned by a committee and the analyst has no authority to change a definition, which makes the trend line they are accountable for permanently unstable.
Remote is normal here and worth using. The work is data, transcripts and written argument, with a periodic need to sit with instructional designers and with whoever owns the operational systems. Fully distributed hires are common; on-premise expectations show up mainly where employee data cannot leave a controlled environment, which is a real constraint in regulated sectors and should be stated in the posting rather than discovered at offer stage. Where headcount planning and skills supply are the same conversation, this analyst works closely with hybrid workforce planning, and the two roles should not be merged into one job description just because both touch skills data.
Common questions
How do I become a Learning Data Analyst?
Pick a real training program and measure whether it changed behavior, using data outside the learning platform. That single artifact beats any certificate. Build the joins yourself, document the confounds honestly, and write up a null result if that is what you find. Add working SQL, comfort with transcript-level qualitative analysis, and enough psychometrics to know why a short quiz does not measure a skill. Institutional research, program evaluation, contact center quality and assessment work are all viable entry paths. Learn one skills taxonomy standard well enough to argue about its levels.
What is the difference between learning analytics and people analytics?
People analytics covers the whole employee lifecycle: hiring, retention, engagement, mobility. Learning analytics is a narrower and deeper slice focused on whether instruction changed capability and behavior, working with tutor transcripts, assessment attempts and course events that a generalist rarely touches. The two overlap at the skills taxonomy, which both depend on. In small organizations one person does both. Above roughly a few thousand employees the learning side usually needs its own analyst, because the transcript-level work does not fit alongside a headcount forecasting cycle.
Can an existing BI analyst do this job?
Sometimes, with a deliberate ramp. A BI analyst brings the pipelines and the modeling, and lacks two things: instructional judgment about what a given assessment can and cannot claim, and experience reading learning interactions qualitatively. Both are learnable in a quarter if the person is curious about the domain. What does not transfer easily is comfort publishing a null result about a program a senior sponsor loves, so check that before promoting internally rather than after.
How do you measure whether AI training actually worked?
Define the behavior first, then find where it is already recorded outside the learning platform: resolution times, rework rates, error counts, ramp-to-productivity, quality review scores. Establish a baseline before the program runs. Compare against a group that did not take it, or at minimum against the same population's prior period, and name the confounds you cannot rule out. Expect selection effects, because motivated people enroll first and were already performing better. A finding you can defend under that scrutiny is worth more than a satisfaction score.
Who should own the skills taxonomy?
One accountable person, with a documented change process and a version history. The taxonomy fails when a committee edits definitions between quarters and nobody records what changed, because every trend line built on it becomes unreadable. The Learning Data Analyst is a reasonable owner when they have authority to accept and reject additions. If ownership sits elsewhere, the analyst still needs veto rights over changes that break comparability, and that should be written into the role rather than assumed.
References
- 1. The Future of Jobs Report 2025 weforum.org About 40 percent of core job skills are expected to change by 2030, and 63 percent of employers name the skills gap as the biggest barrier to business transformation.
- 2. Most In-Demand Jobs In eLearning And EdTech ✓ jobs.elearningindustry.com Data-focused and AI-enabled roles are listed among the fastest-expanding openings across learning and edtech employers.
- 3. L&D Skills Of The Future ✓ thirst.io Data capability and AI literacy appear on the practitioner list of essential L&D skills for 2026.
3 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.