Example · Data Analyst
A model can run the pull. You are the one who knows the table double-counts.
23 of your 40 hours still need you
- 11 stay yours
- 12 faster with AI
- 17 tools could take
A typical week at the middle pace. Your plan uses your own tasks.
Is data analyst a dying career?
By our task-by-task count, 23 of a typical 40-hour week as a data analyst still need a person. Tools could already take about 17 hours. Pulls, deck refreshes, dashboard builds and clean-up are a large part of your week, and a frontier model with warehouse access does that work today. That is the routine work analysts tend to dread.
Read more
What it cannot do is know that a table double-counts something, decide whether a test can support the claim a PM (product manager) wants, or tell a VP that two percent is noise. That is what people come to you for.
This helps you, and it also creates a problem. The exposed work is why a company hires a junior analyst. The work that holds up is work you have grown into and are not yet paid for. If a seat on your team has closed unfilled, yours is safer than that one was. The plan helps you make your job the second kind while you have time.
The hard truth
Nearly half of analyst postings now ask about AI.
Tech hiring is thinner than it was before 2020, and the data postings that remain have changed. AI now shows up in close to half of them.
Sources (2)
- About 45% of US data and analytics postings mentioned AI in December 2025, the highest share of any field. Indeed Hiring Lab, January 2026 Labor Market Update, 2026-01
- Total US tech postings were about 34% below February 2020 levels at the end of 2025. Indeed Hiring Lab, January 2026 Labor Market Update, 2026-01
What stays yours
Knowing which table double-counts what, and why attribution never matches finance
This rests on the cost of checking and on context. A model querying the warehouse returns a number that looks right. Discovering it is wrong takes someone who has been burned by that table before. This knowledge lives in you rather than in the dbt docs, and it is your best asset.
Deciding whether a test can support the claim
The cost of an unchecked wrong answer is highest here. A wrong readout ships a change that underperforms for a year, and nobody finds out. That is why a human who can spot a sample-ratio mismatch or a peeked result is protected.
Turning "can you pull X" into "what would you do differently if X were higher?"
This depends on context. The model answers the question asked. Knowing which question is worth asking requires knowing the team's plan for next quarter and what happened the last time they made that call.
Being the person who says "this number moved, and it isn't real"
This needs accountability and presence. Someone must be answerable in the room when a VP acts on a number. Often that person is you, after the fact. The moves above change when.
Reconciling numbers that disagree
What protects this is context. The model can diff two queries. Knowing that finance books returns on refund date and marketing on order date, and which one the board deck uses, is institutional memory.
What's next after data analyst?
Plan A is your best move inside the work you do now. Most sessions in the 13-week plan build toward it.
- A
Analyst embedded with one team
Why it fits
Why you. You already scope tickets for the teams you serve, and you know their data better than they do. If "in the room when the decision gets made" is where you want to be, this is the move.
The gap. You have no defined role and no recommendations with tracked outcomes yet. Setting the agenda is judged on outcomes and politics, which is different from a ticket count.
The trade-off. You depend on one sponsor; if that VP leaves, a reorg dissolves the seat back into the queue. The queue also protected you, since nobody questions an analyst with 40 tickets. You trade a visible workload for a visible opinion, which can be wrong in public.
Other moves that fit
These are also good bets in this field. Read them if Plan A doesn't suit you.
Own test review: whether a test can support the claim
Why it fits
Why you. If you have ever redone the statistics on a model-written readout, that is the job. Many companies run a few tests a month with no registry or checklist, and PMs read their own results.
The gap. You need to learn power, sample-ratio mismatch, peeking, variance reduction and multiple comparisons. Kohavi, Tang and Xu's book covers all of it. If no registry exists, you build it.
The trade-off. At a mid-size company this is a half-role, with too few tests for a full job, so it only works paired with working embedded with one team. It also means telling product managers "no" in writing, again and again, and being the reviewer who delayed the launch.
Own the metric definitions with the analytics engineer (the backup plan)
Why it fits
Why you. Every quirk you know is a missing test or a wrong model in dbt, the data modeling tool, and the analytics engineer may not know which numbers the board deck uses.
The gap. You need dbt, git and testing habits, plus six months of hands-on work before you can sit the certification.
The trade-off. Tools are taking over the modeling half (Spider 2.0-DBT at 66% and rising, XLang Lab, September 2026), and only the definition half holds. There is often one seat, already taken, and it pulls you toward engineering when the goal is the room. This is the plan if working with one team and test review both stall by week 8.
What's a plan B for data analysts?
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
AI evaluation specialist
Deciding whether a test can support the claim is the same skill AI teams need to judge model outputs and launches.
Other options
AI value and ROI analyst
Reconciling numbers that disagree prepares you to measure whether AI tools save the time and money promised.
LLM evaluation data scientist
Your SQL and experiment review work are a base for LLM evaluation, though most postings want a few more years of data science.
In your own plan you can switch to any of the 47 roles.
The first two weeks
The first 2 of 13 weeks. Most weeks have three sessions of 30, 60 and 60 to 90 minutes, plus one 40-minute Plan B session (45 minutes in week 1). Weeks 11 to 13 have two regular sessions to leave room for planning season. By week 13 you'll have a one-page metrics memo that replaced the weekly deck, a test checklist the team uses before results go to leadership, and a two-page recommendation for a VP.
- Week 1: Measure your week
What you do
- AKeep a time log for one week. Sort your hours by the 14 tasks on this page.
- AExport the last 90 days of data request tickets. Add a column that asks "did the answer change a decision?" and fill it in honestly.
- AWrite down every data quirk you carry in your head, one line each.
- BTake AI Fluency: Framework & Foundations, a free Anthropic course. Then ask AI three questions about a public Census table and check every number yourself.
- Week 2: Test AI on the weekly deck
What you do
- AAsk your manager which AI tools may read the full data warehouse, not just one slice of it.
- AIf the answer allows it, have the approved tool build this week's metrics deck from your BI tool's exports. Time it.
- AList every number the AI got wrong or could not have known.
- BRead Anthropic's guide to defining success criteria. Write criteria for an AI tool that answers one-off data questions.
Three courses from the plan
Checked 2026-09-22.
- Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing (Kohavi, Tang, Xu)Cambridge University Press, 2020 · Price varies by retailer
Yes, Weeks 4–7. It is the single highest-value purchase in this report.
- Accelerating Innovation with A/B TestingMaven, taught by Ron Kohavi · $1,999
Only on company money, and only after “Own test review: whether a test can support the claim” has a named owner. The book covers most of it.
- The Effect: An Introduction to Research Design and Causality (Huntington-Klein)Free online at theeffectbook.net; print via Chapman & Hall · Free
Yes, months 4–6, for the "why did it drop" analyses that aren't experiments.
Risks beyond AI: moderateTech-cycle and supply risk
Analytics teams, especially at consumer and e-commerce companies, grew fast in 2020–21 and were cut in 2022–23 with the rest of tech. Hiring is soft, and entry-level jobs are crowded with certificate holders. Some of what you see comes from that cycle rather than the tool.
Where this could be wrong
- The shares are guesses until Week 1. If your log shows 30% pulls rather than 14%, the share of your week tools can do reaches the mid-50s percent, pressure on entry-level jobs crosses 40 out of 100, and working embedded with one team is urgent.
- Our view that AI handles ad-hoc pulls well rests on a benchmark. Spider 2.0-Snow at 96% is a leaderboard score on a fixed task set, not your warehouse with its undocumented quirks. Being "confidently wrong on our schema" is the counter-evidence.
- We assumed a central team with an intake board. If you are already embedded, the embedded move is done and the plan starts at Week 4; if analytics reports to finance, accountability rises and the report is too pessimistic.
- The 66% master's-or-PhD figure is for the "data scientist" title on Glassdoor US. Product analyst and "decision scientist" roles at consumer companies often skip the degree. We decline the title. The statistics are still worth learning.
Also matched: data analyst, analytics analyst, business analyst. Research checked 2026-09-22. Page updated .