What AI is doing to Data Science jobs

The roles I am best qualified for are the ones AI is most likely to erode. The roles growing fastest are the ones I am least prepared for. I only learned this by pointing the tools of my own trade at my own field.

For the past five years I have built data pipelines and produced analytics across several fields — workforce data, healthcare pathology, and financial-services research. Here I use them to measure, rather than guess, how generative AI is changing what employers ask data professionals to do — across six occupations, and for my own career.

• How the six Data Science roles differ in skill demand
• How demand changed, 2021 to 2026
• How much AI has entered each role
• What this means for my own career

A companion methods post documents the pipeline, the taxonomy, and the statistics in full. This post stays with the findings.

The data

The 2026 figures come from LinkedIn job ads posted in July–August 2026; the 2021 figures from LinkedIn ads posted in September 2021. A language model reads each ad and extracts the skills it requires, and a second step maps every raw phrase (Python, Py3) onto one of about 11,300 canonical concepts — the unit of analysis here. The 2026 sample is roughly 500 ads per role (Analytics Engineer 341); the three roles with 2021 data have 100–150 ads each. Only Data Analyst, Data Engineer, and Data Scientist had enough 2021 ads to compare over time, so every over-time result is limited to those three.

How the six Data Science roles differ in skill demand

“Data Science” is one label for six different jobs. The clearest way to see the difference is each role’s skill-demand profile — how often its ads ask for each family of skills. In the charts below, each axis is a skill category, and the distance along it is the share of the role’s 2026 ads asking for any skill in it. The dashed grey shape is the median across all six roles, so each role is shown against the typical role. There are 27 categories in all; the radar shows the eight that most distinguish the roles, and the drop-downs below give the numbers and examples.

Two groups stand out. AI Engineer and Machine Learning Engineer stretch toward machine learning, foundation models, model deployment, and cloud platforms. Data Analyst and Analytics Engineer stretch toward data transformation and visualisation, and barely register on the modelling axes. Data Engineer and Data Scientist overlap with both groups — which is why the single label hides so much.

To explore this landscape rather than only summarise it, I also built an interactive knowledge graph of the full skill set. It is a supplementary tool: its value is visual and exploratory. Every skill is a node, sized by demand; an edge joins two skills that often appear in the same ad. You can pick a role and see which skills cluster together — something a static chart cannot show.

How demand changed, 2021 to 2026

For the three roles I can track over time, the skill mix has moved sharply. All three show the same broad pattern — legacy tools out, modern and AI tools in — but the specifics reveal how differently AI is arriving in each.

Data Engineer shows the clearest structural change — a move from the old big-data stack to what the industry calls the modern data stack. Java fell from 34% of ads to 12%, and Hadoop from 26% to 7%. In their place, dbt rose from under 2% to 21%, data governance more than quadrupled to 35%, and CI/CD went from 8% to 28%. This matches a well-documented migration: on-premise Hadoop clusters proved costly to maintain, and cloud warehouses plus SQL transformation tools replaced them — a shift Databricks flagged as early as 2021. The surge in data governance and data quality is not incidental: dbt Labs’ 2026 State of Analytics Engineering report argues that AI now produces analytics faster than the governance and trust mechanisms meant to check it, which pushes quality and governance to the centre of the role.

Data Scientist shows generative AI arriving directly. R fell from 63% of ads to 37%, and older infrastructure — Hadoop, Scala, Keras — declined, while large language models rose from 0% to 18% and generative AI from 0% to 15%. This is the industry-wide move from an R-centred toolkit to a Python-and-cloud one: Python is now used by around 58% of developers against under 5% for R in the Stack Overflow Developer Survey. (Note the falling skill here is R and legacy infrastructure, not statistics itself — classical statistics falls in the Data Analyst ads, not the Data Scientist ones.)

Data Analyst changed least, but in a revealing direction: classical statistics (21% to 7%), R (33% to 22%), and data mining all fell, while Power BI rose (32% to 42%). The role is consolidating around business-intelligence tooling and moving away from hand-written statistical work — the kind of routine task automated tools now handle.

How much AI has entered each role

The central question is how far AI skills have actually entered these jobs. They concentrate heavily in the two AI-native roles, as the numbers make plain.

One clarification matters. When an ad lists “large language models” or “generative AI” as a required skill, it is almost never asking for academic familiarity. In context these ads describe applied work. Three real examples from the 2026 dataset:

“Hands-on production experience with LLMs: prompt engineering, agent development, evals, RAG, and deployment at scale”

“Experience building and deploying generative AI solutions using large language models (LLMs) into Production”

“Experience building applications with LLMs, including areas such as RAG, prompt engineering, agentic workflows, evaluation, or fine-tuning”

So “AI skill” in this data means the ability to build with AI, not merely to understand it.

The clearest signal is in the Data Scientist role. Of the 2,802 skills that appear in Data Scientist ads, machine learning ranks 3rd, large language models rank 12th, and generative AI 20th — ahead of many long-standing techniques. Neither LLMs nor generative AI appeared in a single 2021 Data Scientist ad; in 2026 they are in about one in six.

That demand is not limited to technology companies. Of the 500 Data Scientist ads, 25% require at least one AI skill, and the employers asking include mainstream enterprises — Ford, Walmart, Capital One, BMO, Amazon, Intuit — alongside AI-first firms such as Tempus AI. In other words, AI-skill demand in Data Science has already spread beyond the technology sector. (This is a small, single-source sample and company name is only a rough proxy for sector, so I read it as indicative, not definitive.)

To see how these AI skills connect to the rest of each role’s skill set — which skills they appear alongside, and how demand ranks — explore the interactive skill-demand graph.

What this means for my own career

Here the analysis turns on me. I scored my own level on the top skills for each role — 0 (aware), 1 (developing), 2 (advanced) — using my resume and my work-history and portfolio projects as evidence, then compared that against real 2026 demand.

This is not a ranking of one “best” job. A single score would hide what I want to do and how exposed each role is to AI. So I look at two things: how well my current skills match each role, and how well each role fits where I want to go.

To make that concrete: for Data Analyst, the most-demanded 2026 skills are ones I have used for years — SQL, Python, Power BI, data visualisation — alongside a few I am still developing and a few I barely know (Tableau, R, Snowflake). Weighted by how often the role’s ads demand each one, that comes to the roughly 80% shown.

The uncomfortable pattern is that I match best exactly the roles most exposed to AI, and worst the AI-native roles that are growing fastest. My strongest current fit (Data Analyst, Analytics Engineer, Data Engineer) and the direction the field is heading (the AI-native roles) point opposite ways.

That is worth taking seriously. The research points in two directions at once, and both bear on the traditional-versus-AI-role split above:

• The traditional roles I match are the exposed ones. Eloundou and colleagues, in “GPTs are GPTs” (Science, 2024), name data analysis, writing, and programming among the work most exposed to large language models — the exact skill set of the analyst and engineer roles where my match is highest.
• But AI mostly assists rather than replaces — for now. Anthropic’s Economic Index finds AI use is about 57% assistance (a person working with the model) versus 43% full automation. In the near term the person who directs the tool keeps the value, which is an argument for moving toward the AI-native work, not away from it.
• Yet the AI-native roles are the least stable to skill up in. Technical skills now have a half-life of roughly 2.5 to 5 years, and AI tooling changes fastest of all (skill-decay analysis). The specific AI frameworks I learn this year are among the most likely to be obsolete next year.

Put plainly: being hard to replace is less about which of the six job titles I pick, and more about how I work and how fast I keep learning. Work alongside AI. Keep a base of skills that do not decay quickly — analytical thinking, judgement, and communication, which the World Economic Forum’s Future of Jobs 2025 report names among the most valued and fastest-rising human skills (it expects about 39% of core skills to change by 2030). One honest caveat: my own ad data cannot confirm that claim — job ads rarely spell out human skills (communication appears in only 10–15% of ads, analytical thinking in almost none), which is a real limit of ad-based analysis. So for that point I rely on external research, not my data.

My learning plan

Three directions, each tied to what the data above shows.

1. Build with AI, and learn the methods rather than the tools. The AI-native roles are where I match worst, and their ads (quoted above) ask for RAG, evaluation, and orchestration repeatedly. Those are methods that outlast the specific frameworks — and the skill-decay research is the reason to learn the method and treat this year’s library as replaceable. My anchor is extending the project behind this post into an AI-served feature.
2. Build things people use. This is my strongest cluster already (the ~68% match on Data and Analytics Engineer), and the rising skills there are learnable and durable: dbt (up from 2% to 21% of Data Engineer ads), a cloud warehouse (Snowflake and Databricks both rising), and shipping one real analytics or education application of my own.
3. Invest in the durable human skills. Job ads barely name them, and my data cannot measure them — but that is rather the point: knowing where value lies, communicating it, and deep domain knowledge are hard to automate precisely because they are hard to specify. It is the least certain of the three bets, and the one I would stake the most on.

How this was built

The pipeline and the method are fully reproducible: raw job ads are turned into typed, categorised, and facetted skill concepts, then demand, change over time, and skill co-occurrence are computed and served from a small BigQuery warehouse. The exact figures depend on the specific job-ad sample, which I do not republish, so the numbers here are not meant to be re-derived by a reader — but the taxonomy, the statistics, and the design decisions are all documented in the methods post.

Thanks for reading. The interactive graph is here; questions and corrections are welcome.