How to Hire a Data Scientist: Sourcing, Screening & Salary Benchmarks

Most “data scientist” reqs we see on intake calls are not the role the company actually needs. About half describe a product analyst with SQL. About a quarter describe an ML engineer in disguise. The remaining quarter, the real ones, are the hardest searches we run, and they cost $112,000 to $200,000 base for the people you actually want. The job title hides three specializations. The resume layer hides almost everything else, ever since the LLM-resume era broke surface-level filters about two years ago.

$112,590. That is the BLS median for the role as of May 2024, repeated by every salary article on the search results page, and it is the number that has gotten more hiring managers into trouble in our queue than any other figure in the last year. The people you actually want to hire do not sit anywhere near it.

Three weeks ago I had a four-minute intake call with a head of growth at a Series B fintech in Irvine. She wanted “a senior data scientist.” Her budget was $145K. She wanted that person to own the experimentation platform, run weekly readouts to product, and “find some patterns in the user behavior data we’ve been collecting for the last eighteen months.” I asked one question. Did she already have a warehouse, or was the user behavior data sitting in PostgreSQL replicas and a Mixpanel export. Mixpanel and Postgres. No warehouse. No analytics engineer. One Looker dashboard somebody had built last year and never updated.

What she actually needed was an analytics engineer and a product analyst, in that order, for less money than her data scientist budget, and the model work she was hoping to enable later would have to wait six months for that foundation. I told her so. She paused, said “huh,” and asked me to bid the analyst search instead. Different role, smaller fee for us, much faster fill, dramatically better outcome for her. That conversation happens four or five times a month.

I’m Gregg Flecke. I run business development at KORE1, which means I spend my days on intake calls with people who think they need to hire a data scientist. The list of times I have talked clients out of hiring through us is longer than the list of times I closed them. That sounds bad for business and it isn’t. Clients who get the role right come back. Clients who let us push them into the wrong hire do not. So when I tell you the version of this conversation we have on intake calls, the bias I have is toward you not paying us for a search that ends in regret.

Everything below comes from the data scientists we have actually placed through our data science and data engineering staffing practice and the searches we have run over the last eighteen months. Some of it will not be flattering to staffing firms, including ours.

What Hiring a Data Scientist Actually Means in 2026

A data scientist applies statistics, experimentation, and machine learning to messy business data in order to answer questions a leadership team cares about and cannot answer with a dashboard. The role is distinct from analyst, data engineer, and ML engineer, and the Bureau of Labor Statistics groups all of them under different occupational codes for a reason. The work is closer to applied research than software engineering on most days.

The market context matters more than the salary numbers. The Bureau of Labor Statistics puts projected growth for the role at 34 percent from 2024 through 2034, with roughly 23,400 openings a year over that decade and a May 2024 median wage of $112,590. Honest numbers. Also the numbers that get every hiring manager into the wrong ballpark, because the median includes a long tail of “data scientist” titles at insurance carriers and government agencies that do work closer to traditional analytics. The people writing the JD that arrives in our queue are usually trying to hire from the upper tier of that distribution at the salary the median implies. The math does not work.

What is happening underneath the BLS curve is harder to see in a single statistic. The January 2026 Data Science Job Market Report from Interview Query describes the current market as a stabilization phase, not a sharp rebound. Causal inference skill mentions are up 17 percentage points year over year. A/B testing is up 14. Pure modeling roles are flat. The story the data tells, if you read past the top-line “data scientist demand is strong” headline, is that demand for the experimentation-and-causal version of the role is climbing while demand for the let-us-build-models-from-scratch version is not. Understand which version you are hiring for and you avoid most of the pain that follows. Skip that step and you will spend ninety days screening for the wrong skill set.

For the role comparison and the question of whether you actually want a scientist or an engineer first, Devin wrote the data engineer vs data scientist breakdown earlier this month and it is the right place to start. Read that one first if you are not yet sure which side of the line your problem sits on. I will not retread it here.

The Three Data Scientists Hiring Managers Confuse

Walk into any “we need a data scientist” intake meeting and ask the hiring manager what the person will spend the majority of their time on. You will get one of three answers. They correspond to three different jobs that the market lumps together for keyword reasons.

The Product Data Scientist

This is the most common version of the role being hired in 2026 and the one most companies actually need. They run experiments on the product, build and maintain the experimentation platform if one exists, work next to product managers on metric definitions, and answer questions like “did the new onboarding flow move retention or are we just measuring a seasonal bump.” They write SQL all day. They write Python in notebooks for analysis, not for production. The artifact at the end of their week is usually a memo, not a model.

Levels.fyi puts the median total compensation for a data scientist at $175,000 and the median for product DS at major tech companies sits closer to $230K once equity is included. Glassdoor’s broader sample, which includes mid-market and non-tech companies, lands around $154,807 in average total pay. Both numbers describe the same role at different parts of the market. The screening question that catches a fake product DS in fifteen minutes is this one. Tell me about an experiment you ran where the result was statistically significant but you recommended against shipping anyway. A real product DS has at least two of those stories. A fake product DS will tell you about a model they built.

The Decision and Causal Inference Scientist

Rarer. Harder hire. Usually has a graduate degree in something with statistics in the name, often economics or biostatistics rather than computer science. Their job is to answer questions where you cannot run an A/B test, which turns out to be most of the questions executives actually care about. Did the marketing campaign cause the lift, or did the lift cause the marketing decision. What would have happened if we had not acquired that company. Should we change the pricing tier and what is the counterfactual.

These hires take longer to fill because there are fewer of them and the ones who exist tend to stay where they are. Senior causal inference talent runs $180K to $260K base in the markets we cover, plus equity at the higher end. The screening signal is whether they can describe a real causal study they have run, what they assumed, and what they would not believe about the result. If they reach for “we’d run a regression” within thirty seconds, they are not the person.

The ML Modeling and Applied Science Hire

This one has been drifting toward ML engineering for four years and the drift is now mostly complete. If your team needs someone to fine-tune a foundation model, build a recommendation system, or own a production model serving pipeline, you are not actually hiring a data scientist anymore. You are hiring an applied scientist or ML engineer who happens to use a data scientist title because the JD template was already written. We send those searches to our AI and ML engineering hires recruiters because the candidate market, the interview loop, and the salary band are all closer to the engineering side of the house. Pay is wider here than the other two specializations, anywhere from $160K to north of $300K base for the people Meta and Google are bidding on.

One thing none of the three specializations need is what most JDs put in the first bullet point. Ten years of “end to end ML lifecycle experience.” That phrase is a tell. It usually means the person who wrote the JD copied it from another company that copied it from a third company in 2019, and nobody has updated it since. The phrase does not appear in the resumes of the strongest data scientists we have placed in the past year. Read what that means.

What Data Scientists Actually Cost in 2026

Salary data on this role is messier than salary data on most engineering roles, partly because the title covers three jobs and partly because the aggregators sample different populations. Here is what the major sources show in early 2026. The disagreements between them are real and informative.

Reading that table requires one note. BLS reports a single median across the entire occupation, so the number sits low compared to what you will actually pay for the people in our queue. Glassdoor’s average runs above the BLS median because Glassdoor’s sample skews toward people who self-report on Glassdoor, which skews toward tech-employed urban data scientists. Levels.fyi runs higher still because their sample is dominated by FAANG and FAANG-adjacent. The honest read is that BLS is the floor for the role as government statisticians count it, Levels.fyi is the ceiling for top-tier tech, and your company sits somewhere on that spectrum based on industry, geography, and equity offering.

Geography moves these numbers more than most articles admit. New York and San Francisco both run roughly 30 to 35 percent above the national median for mid-level data scientists. Seattle adds about 25. Austin compressed in the last eighteen months and now runs maybe 8 percent above national, down from 20. Orange County, where most of our placements happen, sits about 10 to 15 percent above national for product DS, more for causal hires because the supply is so thin. Our salary benchmark tool can give you a tighter number for your specific market and specialization.

For contract rates, take the all-in annual you would offer a direct hire, multiply by 1.4, and divide by 2,080 hours. A $170,000 senior product DS comes out to roughly $114 per hour fully loaded through a staffing firm. That number includes our margin, the contractor’s burden, and the flexibility of being able to end the engagement at any time. A surprising number of clients now choose contract data scientist arrangements for the first six months and convert if it works, especially in markets where they are not sure the role is the right shape yet.

Sourcing in the LLM-Resume Era

LinkedIn keyword search broke for this role about eighteen months ago, and the reason is something every recruiter on our team complains about constantly. Every data scientist resume on the platform now lists Python, SQL, scikit-learn, PyTorch, TensorFlow, A/B testing, causal inference, dbt, Snowflake, and the other twelve keywords an LLM will inject when asked to “optimize my resume for the job description.” The boolean search returns ten thousand profiles that all look identical. The actual quality differentiation happens below the keyword layer and the platform’s filters cannot see it.

What our recruiters do instead is unsexy and slower. They start from portfolio first. A real product DS has a personal site or a Substack with at least one writeup of an experiment they ran, what they expected, what surprised them, and what they would do differently. A real causal scientist has a notebook somewhere, often on GitHub, that walks through a study with the assumptions made explicit. A real applied scientist has commits to an open-source ML library or a public Kaggle competition history that shows progression rather than a single notebook submission. None of those signals show up in the keyword string. All of them take a recruiter five minutes to verify.

The other thing that changed. GitHub star count is now a worse signal than commit pattern. A data scientist with four hundred stars on a single repo from 2022 and no commits since is not active. A data scientist with thirty stars across nine repos and weekly commits is active. The platform’s “rising star” feature ranks the wrong one higher.

If you are running this search yourself and you have a sourcer who is good at the resume layer but not at portfolio audit, you will make the same hire your competitors made, and you will pay the market clearing price for it. If you want to talk to the people whose portfolios actually demonstrate the work, you either need a sourcer who has been doing data science specifically for at least two years, or you need to bring in help. Our practice is one option. There are others. The honest version is that any specialized firm will outperform a generalist agency by a wide margin on this specific role.

How to Screen Without Wasting Five Hours on a Take-Home

The standard data science take-home in 2026 is broken in both directions. It is too long for strong candidates, who turn it down because they have three other offers and your test costs them a Sunday. And it is too short to actually filter, because the work it asks for, “predict churn from this CSV in a notebook,” is exactly the work an LLM does well. You will get back beautiful notebooks that all look the same, and you will not know which candidate wrote which paragraph.

Replace it with a 45-minute structured conversation that does three things. Give the candidate a real ambiguous business question from your company, blinded as needed, and ask them how they would approach it. Listen for clarifying questions. The strong candidates ask three to five before they start sketching an approach. The weak ones start sketching immediately. Then do a SQL whiteboard that is not a leetcode problem, with a small schema and a question that requires a window function and a thoughtful join, and see whether they can talk through their reasoning while they write. Last, hand them a deliberately flawed analysis, two pages or less, and ask them to critique it. That is the best signal in the entire interview loop. A senior data scientist will find three problems in five minutes. A junior one will find one and spend the rest of the time praising the approach.

That whole conversation takes less time than scoring a bad take-home. It tells you more. And it does not lose you the strong candidates who refuse the take-home in the first place.

When to Hire a Data Engineer or Analyst Instead

A short section because most of this argument lives in other posts.

If you do not have a data warehouse yet, and your data is sitting in production database replicas and SaaS exports, your first hire is not a data scientist. It is a data engineer. The complete version of that argument lives in how to hire a data engineer instead, and the short version is that a data scientist hired into that environment will spend nine months building the warehouse instead of doing the work you wanted them to do. That is the most common painful mistake we see.

If your actual need is dashboards, weekly reporting, and “tell me what happened last quarter,” you need an analyst, not a scientist. The cost difference is substantial and the screen is different. We covered that in hiring a data analyst in March.

If your need is closer to fine-tuning foundation models or owning production model serving infrastructure, that is an ML engineer or applied scientist hire. Different talent pool. Different interview loop. Different bands. Start at the AI and ML engineering hires guide linked above.

The reason I bring all three up in a guide that is supposed to be selling you on a data scientist hire is that getting the role right matters more than which firm runs the search. A data scientist hired into the wrong shape of problem is one of the most expensive mistakes in hiring, because the cost is invisible until quarter three, and by then the person has built relationships and you do not want to let them go.

Common Questions Hiring Managers Ask Before Hiring

How long does this search realistically take?

Forty-eight to ninety days for a senior product DS through a specialized firm, longer if you are running it in-house from a cold start. The causal inference specialization runs longer because the candidate pool is thinner. ML modeling sits closer to the engineering search timeline, sixty to a hundred days for the right person. None of that includes your interview loop, which is the part most clients underestimate.

Can a data engineer grow into a data scientist?

Wrong direction, usually. The skills overlap on Python and SQL and almost nowhere else. A data engineer trying to grow into a product DS role is essentially starting over on experimentation, statistics, and product judgment, and most of them find they do not want the work after six weeks. The opposite move, scientist to engineer, happens more often and works better.

PhD or no PhD?

Depends on the specialization. Product DS roles do not require one and the strongest product DS hires we have placed do not have one. Causal inference is the place where graduate training in statistics, economics, or biostatistics still matters, and a PhD is a real signal there because the work requires comfort with assumptions most people skip past. ML modeling is mixed. A PhD in machine learning helps for foundation-model work and helps much less for everything else.

Should we hire one full-time or contract first?

Contract-first is more common in 2026 than it was two years ago, especially for the first data scientist on a team. The reason is that the wrong shape of role is the most common failure mode, and a contract engagement gives you ninety days to find out which version of the role you actually need before you commit a permanent slot to it. We see roughly a third of our data scientist placements start as contract or contract-to-hire now, up from maybe one in ten three years ago.

What is a fair take-home test length?

Two hours maximum, or skip it. Anything longer screens out the candidates you most want to hire, who have other offers and will not spend a Sunday on your test. The forty-five minute structured conversation described earlier is a better filter than a four-hour notebook anyway.

Do you actually need a data scientist, or are you describing an analyst?

I ask this on every intake call, and the answer is “an analyst” about half the time. If the work you want done is build dashboards, run weekly reporting, answer ad-hoc questions about last quarter’s numbers, you are describing an analyst. If the work is design experiments, build predictive models, do statistical inference on messy data, you are describing a scientist. The salary delta between the two is substantial. The wrong call costs you six figures and a year.

If you want a second opinion on which version of the role your company actually needs, or you want to start a search with a recruiting team that has run this specific search hundreds of times, talk to our data team. We will be straight with you about whether we are the right fit, including the times we are not.