Example · Data Scientist
One prompt now fits, tunes and scores a model. Saying whether the test can support the launch is still on you.
23 of your 40 hours still need you
- 15 stay yours
- 8 faster with AI
- 17 tools could take
A typical week at the middle pace. Your plan uses your own tasks.
Is data science a dying field?
By our task-by-task count, 23 of a typical 40-hour week as a data scientist still need a person. Tools could already take about 17 hours. Fitting and tuning standard models got cheap this year. AI agents now win medals in most Kaggle-style contests, and data platforms ship agents that profile, split, tune and report a model from one sentence.
Read more
That is the part of the job most people were hired for, and it is the most exposed. About 43% of your week sits there.
What holds is the work that fails without anyone noticing: experiment design, causal claims, model validation, and the newer job of proving an AI feature is good enough to ship. Those are expensive to check and depend on facts nobody wrote down. The plan moves you upstream, from reading tests after launch to reviewing them before, and builds a record of what the tools got wrong.
The hard truth
Most of the models you build may never ship.
A lot of data science work ends at the proof of concept. Companies scrap close to half of AI pilots before production, and more are abandoning most of their AI projects.
Sources (2)
- 42% of companies abandoned most of their AI initiatives in 2025, up from 17% in 2024. S&P Global Market Intelligence, via CIO Dive, 2025-03
- The average organization scrapped 46% of AI proofs of concept before they reached production. S&P Global Market Intelligence, via CIO Dive, 2025-03
What stays yours
Designing experiments: randomization unit, power, guardrails, stopping rules
This is hard to verify. A bad design is discovered weeks later, after the launch decision, and the cost is a wrong product call nobody reverses.
Answering causal questions from observational data
This rests on context. Whether a launch caused the lift depends on how the data was generated: a promo nobody logged, a region that rolled out early. Those facts live in people's heads.
Validating and monitoring models for leakage and drift
This is hard to verify. A leaked feature looks like a great model. Catching it means knowing which columns could not exist at prediction time.
Evaluating AI and LLM (large language model) features
This is hard to verify. A model can grade outputs, but someone has to show the grader agrees with people and that the test set looks like real traffic.
Reconciling numbers that disagree
This depends on context. Nobody has written down which table finance trusts or why the other one double-counts.
What's next after data scientist?
Plan A is your best move inside the work you do now. Most sessions in the 13-week plan build toward it.
- A
Own experiment design and review
Why it fits
Why you. You already run readouts and you already see the bad designs. Being asked too late is the problem, so move the moment you are asked.
The gap. You need a habit of writing test plans in advance, checks for uneven traffic splits and early peeking, and the words to say no early.
The trade-off. You will build fewer models and your output will look smaller on any dashboard of shipped work. Some managers reward volume.
Other moves that fit
These are also good bets in this field. Read them if Plan A doesn't suit you.
Move into AI and LLM evaluation
Why it fits
Why you. Statistics on noisy labels, agreement and confidence intervals is the core of eval work, and you already distrust a metric until you know how it was made.
The gap. You need a public write-up of an AI test: a test set, a scoring guide, a number for how well the grader agrees with people and one problem you caught.
The trade-off. These postings ask for 3 to 4 years and often a graduate degree. Early on you will likely do this inside your own company first, without the title.
Move into regulated modeling in banking, insurance or health (the hedge)
Why it fits
Why you. These employers need people who can document and defend a model to an independent validator, and your validation habits are the starting point.
The gap. You need practice writing model documentation, testing for fairness and waiting through review cycles.
The trade-off. Slower work and slower releases, and pay can trail tech. The guidance scales with bank size, so smaller institutions expect less.
What's a plan B for data scientists?
Plan B is a job AI is creating that these skills already fit. One session a week in the plan trains you for it, so it is ready if your job shrinks or you want a change.
BPlan B, picked for this role
LLM evaluation data scientist
You already put confidence intervals on noisy metrics and distrust a number until you know how it was made, which is the core of judging an AI grader.
Other options
AI evaluation specialist
Your habit of building holdout sets and checking a model against them carries straight into writing test sets and rubrics for AI systems.
AI value and ROI analyst
You design experiments and measure lift, which is exactly how a company should judge whether an AI project paid off.
In your own plan you can switch to any of the 47 roles.
The first two weeks
The first 2 of 13 weeks. Each week has three sessions of 30, 45 and 60 minutes, plus one 40-minute Plan B session. Throughout, you keep a catch log: each time a tool, test or model is wrong, you write down what happened, how you caught it and what it would have cost. By week 13 you'll have a catch log of agent and test errors, a one-page memo asking to own pre-launch review, and a published write-up of your AI testing.
- Week 1: Measure your week
What you do
- ATrack one week of your work against the task list on this page. Write down roughly how many hours went to each task.
- AList your last ten analyses or models. Next to each one, write the decision it fed.
- ARead your company's AI tool policy. Write down which data may go into which tools.
- BTake the first half of Anthropic's free AI Fluency course. Write down what you would hand to an AI and what you would check yourself.
- Week 2: Check an AI agent's work
What you do
- APick one finished analysis whose answer you trust. You will use it as an answer key.
- AGive your approved AI agent the original question and let it run on its own.
- ACompare the agent's answer with yours. Start a catch log: what it got wrong, how you caught it and what it would have cost.
- BFinish the AI Fluency course and pass its final assessment.
Three courses from the plan
Checked 2026-10-15.
- Occupational Outlook Handbook: Data ScientistsBLS · Free
Yes, once. The growth projection is the counterweight to the layoff headlines.
- O*NET 15-2051.00 Data ScientistsO*NET OnLine · Free
Optional. Useful for checking which of your tasks the official description still lists.
- The Effect: An Introduction to Research Design and CausalityNick Huntington-Klein · Free online
Yes, Weeks 5 and 7. It is the best free source for the causal questions that hold.
Risks beyond AI: moderateBudget and layoff-cycle risk
Budget cycles and layoff waves move data science headcount more than any model does. When a company misses its numbers, analytics and research teams are easy to trim, and tech layoffs ran past 108,000 by mid-May 2026. Against that, the government still projects much-faster-than-average growth for the occupation.
Where this could be wrong
- Benchmarks overstate fit to real work. MLE-bench, an AI test built from Kaggle contests, uses clean Kaggle tasks with a fixed metric, while real model work starts with a vague ask and messy joins, so fitting may be less exposed than scored here.
- Databricks' 77.1% internal benchmark for Genie Code has no published method. We used it only as a sign of direction.
- Your company may already have a strong experimentation team. If design review is owned, the experiment design move shrinks and AI evaluation becomes the lead.
- Evaluation work could fold into ML engineering titles, leaving fewer data scientist evaluation seats than the Apple and Perplexity postings suggest.
Also matched: data scientist, junior data scientist, product data scientist. Research checked 2026-10-15. Page updated .