top of page

The Data Scientist is Dead? Long Live Data Science?

23 minutes ago
5 min read

What data folks do once everyone can write code



Data science had a short but bright moment in the spotlight. The nerdy statistician learned to market themselves in the early 2010s, rebranding and updating their skillset to "data scientist". And then some subset of those nerds moved on to "machine learning" and then further rebranded to "machine learning engineer" and "AI engineer". And most recently to "forward deployed engineer" -- updating their paycheck along the way.


Now, they may be a little angsty about their role given that many more individuals are empowered to write code and wield models with the latest agentic tools. Practically anyone now can generate a quick prototype and build an internal coalition around it. In years past, a data scientist that could simply collect and display data in a neat visualization was often sufficient. Not anymore. Anyone with an hour or two to spare can take that same data and fit a naive xgboost classifier. Or even skip the boosted model and ask Chat to classify on the fly.


A whole swath of these data workers recently had to learn the realities of working in the private (and public) sectors in real time -- the value workers provide is constantly evaluated and eventually judged against tangible business outcomes, not the promise of expertise. The marginal value on the Pareto curve that management expects has moved rightward. And moving up that value chain can be difficult. So if anyone can fit a boosted model, what's the role of us, the data folks?


The thread that connects the various data roles over the years is statistical judgment: using the tools at our disposal to question assumptions, catch a mis-framed problem, and implement the appropriate solution. And it now matters most when a prototype starts to catch on. Across the teams we work with, I’ve seen this hold at every point on the data spectrum: visualization, business analytics, prediction models, statistics, neural networks, and now LLM endpoints and pretrained models. The data worker’s next evolution: from building the whole pipeline to pressure-testing data prototypes once they take off.


There is no decision tree


LLMs and agentic tools are of course not going away any time soon. How does the data worker "provide value" if everyone else is wielding the same tools? One of my big takeaways from spending two years getting a (perhaps obsolete?) statistics master's degree is that statistics and statistical thinking are not a directed graph that you feed forward to get to your answer. I remember more than one student asking for a sort of "decision tree" of when to apply which model to what kind of problem. And the resulting look of puzzlement of various professors (who remain unnamed but likely rank somewhere in the top decile of statistics academics). I'm not sure if it was because they have never thought of such a decision tree, that it would be too difficult of a task, or that it just didn't comply with their own mental model of the statistics profession.


I suspect it was the latter. In Richard McElreath's famous courses on Bayesian statistics, he refers to the teachings as "statistical rethinking". That suggests that statistics is a way of approaching the world; not a set of distinct tools you pick up, use, and disregard. The value of the data worker is knowing how to use these tools to question the assumptions, spot when a problem has been framed incorrectly, and rethink the appropriate solution. A fixed, finite decision tree would overfit the problem.


There's a great post from Hamel Husain on the "Revenge of the Data Scientist". They argue, in my over simplified terms, that the toolset of the data scientist -- analysis, metrics, validation, monitoring, etc -- still stands. One inevitable conclusion is that the data science process needs to be co-opted by LLM users to make better decisions. Data science can be a harness in a similar, but more abstract way, like Claude Code is a harness to the underlying LLM. It provides the directions and lane bumpers. So anyone leveraging LLMs to perform data work would be better off by applying a data science worldview, or principles, to it. The role of the data worker today is to use agentic coding as leverage without forgetting about the underlying principles of statistics and data science. Whether it is to translate the viz in your head into ggplot code or to use statistical principles to write a cross validation approach that allows you to test many models -- expedited by Claude Code writing the tidymodels or statsmodels boilerplate.


When the demo takes off


The business user who wields LLMs effectively will build the demo and rally internal interest. In some cases maybe they can insert a new off-the-shelf discriminative model like Jev and get great results fast without the data science harness (we previously blogged about comparing pre-trained to the "old school" approach). The data scientist's moment is when that internal demo takes off but before it hits production. That means working with the business to refine the demo, build the lane bumpers, refine the model, and move it up the Pareto curve from 90% to 99% -- rather than from 0-90% like in years past. 


That last marginal stretch is where LLMs fall short and where a trained data worker can apply their expertise. An ICLR 2026 benchmark built around pitfalls like Simpson’s paradox and selection bias found significant limitations in how LLMs handle statistical causal inference. These are common issues and would be easily caught by someone who knows what they’re looking for.


The job title will change again. We'll get a new fancy label for these specialized data workers (remember prompt engineer?). But of course the underlying fundamentals will be the same: question assumptions, catch a mis-framed problem, and implement the appropriate solution.


Got a demo that is starting to take off?


If a demo like that is making the rounds at your company right now, this is the part we do.


This is the kind of problem we work on at Lander Analytics. We help teams take a prototype someone built in an afternoon and find out whether it holds up: how the problem was framed, how the model was validated, and which method the problem actually calls for. Whether you need a second set of eyes on a demo or help building the version that goes to production, reach out at info@landeranalytics.com.



Joe Marlo

Director of Data Science

Lander Analytics

 

Subscribe to our Substack and below to our monthly emails for practical AI strategies for your organization: what to build, what to avoid, and how to make systems reliable in the real world.


Work with us: If you want help identifying the right first workflow, building a permissioned knowledge base, or training your team to ship responsibly, reach out at info@landeranalytics.com. 


About the author: Joe Marlo is Director of Data Science at Lander Analytics, where he designs agentic workflows, statistical models, and interactive frontends that put rigorous analysis into production.

Get our latest blog posts—delivered monthly!

  • X
  • LinkedIn - White Circle
  • Bluesky
  • Untitled design (53)
  • YouTube - White Circle

© 2026 Lander Analytics

bottom of page