The Data Scientist Who Watched AI Learn Her Job (Then Lost It)

Two Data Scientists, One Moment in History

On May 20, 2026, Moyan Chen came into work at Meta's Instagram team and didn't come back. She was 24, had three big-tech internships behind her, and had spent nearly a year quietly watching AI absorb her job from the inside. "It got to the point where I wouldn't check AI-generated queries because they had gotten so accurate," she told Business Insider. "Ultimately, I lost my job to AI."

She felt more relief than pain. The dread had been building for months. Every Tuesday night she'd leave the office wondering if she'd be back. Every Wednesday morning she'd check her email early.

What Chen was doing all day — validating AI-generated queries, checking outputs, running SQL — was exactly the work the AI learned to do without her.

Now consider Sara Nobrega, also a data scientist, also at a large company, also in 2025. She didn't get a layoff notice. She got uncomfortable. She started asking herself a different set of questions: How will this model actually go live to users? How will it add value? How will it still work in six months? Those questions — not a new certification, not a bootcamp — led her to a new role as an AI engineer within a year, inside the same kind of large-company environment she'd always worked in. Same industry conditions. Different questions. Different outcome.

The gap between these two stories isn't luck or timing. It's what each person was doing with their time — and whose job it was to ask what.

To understand why these two stories diverged, you need to know exactly which tasks AI has already absorbed and which ones it still cannot do without a human holding the wheel.

What AI Has Already Eaten — and What It Hasn't

AI has already automated a specific, nameable set of data science tasks. But its accuracy decays without human maintenance, which means the human role hasn't disappeared — it has moved upstream to the layer the AI cannot govern itself.

Think of your daily work in three tiers.

Tier 1 covers what's already gone or going: data cleaning, preliminary analysis, SQL generation, standard visualization, and the systematic testing of dozens of model configurations. The World Economic Forum estimates that 50 to 60 percent of typical junior data science tasks — data cleaning, coding fixes, research synthesis — are already AI-executable. If you're spending most of your day on these tasks, you're working in the highest-exposure zone.

Tier 2 is the dangerous middle: conditionally automated, but quietly decaying. Anthropic deployed Claude-powered analytics agents that automated 95 percent of business analytics queries at roughly 95 percent accuracy. Impressive — until you look at what happened next. Without active human curation of the underlying reference documentation, accuracy dropped to around 65 percent within a single month. Anthropic's own data scientists didn't disappear after deployment. They shifted their role to maintaining the "skills layer" that keeps the AI from silently degrading. The AI needs a keeper. That keeper is still a human.

Data scientists are the professionals best positioned to identify opportunities for automation, design approaches for testing and monitoring systems at scale.
— Cassie Kozyrkov, Former Chief Decision Scientist at Google

Tier 3 is the defensible core: defining which metrics matter, framing the business question before touching any data, translating model output into a decision a stakeholder can actually act on, and owning adversarial review when AI output goes wrong. As former Google Chief Decision Scientist Cassie Kozyrkov has argued, data scientists are the professionals best positioned to identify where automation belongs — which means the highest-value work is deciding where AI goes, not running it once it's there.

The Tier 2 finding deserves a second look. AI output that looked reliable at launch quietly became unreliable within 30 days. The human who owns governance is not a remnant of the old job — they are the reason the AI keeps working at all.

This three-tier framework holds regardless of your specific domain. Whether you're doing product analytics, marketing attribution, HR people-analytics, or financial modeling, the pattern is the same: AI can generate the first draft, but it cannot own the question it's answering or verify that its answer is still valid next month.

Knowing which tasks are at risk is necessary. But the macro picture matters too, because it tells you whether the profession itself is contracting or simply reshaping around different work.

The Profession Is Reshaping — But Around What?

Moyan Chen represents something BCG named explicitly in its April 2026 study of 165 million U.S. jobs across 1,500 roles. Data and financial analysts landed in the "Substituted Roles" bucket — roughly 12 percent of current jobs where AI productivity gains are likely to reduce headcount rather than drive hiring. In practical terms: a team of five junior analysts might become two senior ones.

The layoff headlines are real. TechCrunch's running tracker counted approximately 120,000 tech roles eliminated in the first half of 2026 that explicitly cited AI — including around 8,000 at Meta, 4,800 at Microsoft, and Salesforce's February 2026 cuts that specifically targeted the data analytics and Agentforce AI departments. These aren't speculation. They're named companies with named executives.

And yet the Bureau of Labor Statistics projects 34 percent employment growth for data scientists from 2024 to 2034 — the fourth-fastest-growing U.S. occupation — with 23,400 average annual openings. A 2026 analysis of 1,000 job postings found entry-level data science salaries at $152,000, up $40,000 from 2025, with senior roles at $215,000 and above. The number of jobs is not collapsing. The per-firm hiring bar is rising.

Both things are true simultaneously. The firms cutting analyst headcount are the ones with mature, production-deployed AI analytics stacks — only about 22 percent of enterprises, per Anaconda's 2026 survey. The other 78 percent haven't yet reached that threshold. The layoffs making headlines are the front edge of a distribution, not the median experience.

BCG's framework is explicit about where the pain concentrates: entry-level and junior positions face measurably higher automation exposure than senior ones. If you're three to five years into a data science career and doing primarily execution work — generating outputs rather than framing questions — you are in the highest-exposure band regardless of industry.

The profession is reshaping, not vanishing. But reshaping toward what? The people who have already navigated this shift point to one consistent pattern in how they moved.

The Three Questions That Tell You Where You Actually Stand

The most useful thing you can do right now is not research new certifications. It's audit last week's actual work against the three tiers, because that audit tells you where you personally sit on the exposure curve.

Sara Nobrega's pivot didn't start with a bootcamp. It started with five diagnostic questions about her existing work. Those questions forced her to distinguish between what she was doing — building models — and what her employer actually needed: models that work reliably in production six months later. That distinction was the pivot.

Having a skill to talk in a nontechnical way is probably the most valuable skill that I bring.
— Milica Cvetkovic, AI Consultant at Google

Max Buckley's path tells a different but complementary story. He spent 13 years taking roughly 40 online courses at Google, moving from financial analyst to leading an LLM information-retrieval research team. He described the process as unstressful precisely because he wasn't racing a deadline — he was compounding a direction. The timeline matters less than having one.

Here's the audit. Three questions, answerable from memory about last week.

Question 1: How many of your tasks involved generating the first version of something — a query, a report, a model? These are Tier 1 tasks. Highest AI exposure.

Question 2: How many involved verifying, correcting, or contextualizing AI-generated output? These are Tier 2 tasks — valuable now, but only if you own the governance layer, not just the checking behavior.

Question 3: How many involved deciding whether a question was worth asking at all — setting a metric, framing the business problem, or translating data into a decision a stakeholder can act on? These are Tier 3 tasks. The defensible core.

The goal isn't to have all Tier 3 work immediately. It's to know your current ratio and deliberately shift it.

Most data scientists, if they're honest, will find their week is 60 to 70 percent Tier 1, 20 to 30 percent Tier 2, and 10 percent or less Tier 3. That ratio isn't a judgment. It's a starting point. The practitioners who navigated the 2025 to 2026 disruption without losing their roles were the ones who shifted that ratio before they were asked to.

This audit works for any data-adjacent role — marketing analyst, HR people-analytics, financial modeler, product analyst. The tier labels don't change by job title. What changes is which specific tasks fall where.

The Sentence That Landed Differently After the Layoff

After losing her job at Meta, Moyan Chen said something that is easier to hear now than it would have been before: "If you only know how to code, that's not enough."

She wasn't talking about learning a new programming language. She was describing the difference between Tier 1 work and Tier 3 work — between generating outputs and owning the questions those outputs are supposed to answer.

The original question — "Will AI take data science jobs?" — is actually the wrong question. AI has already taken specific tasks from data scientists. The profession is reorganizing around the tasks it hasn't taken yet: the metric-setting, the problem-framing, the translation of model output into a decision someone can act on, and the governance layer that keeps AI from quietly degrading on its own. Those are the tasks that kept Sara Nobrega employed. They are also the tasks that Anthropic's own AI agents still require a human to own.

Block 30 minutes this week — before you open any course catalog or job board — and categorize last week's actual tasks into the three tiers. Count how many were Tier 1, Tier 2, and Tier 3. If your Tier 3 share is under 20 percent, pick one specific recurring task where you currently skip straight to the data, and practice articulating the business question out loud first. That habit — not a certification — is where the shift starts.

The question was never whether AI would take data science jobs. The question is which data scientists are doing data science's most human work — and whether you're building toward that, or away from it.

Recommended Tools & Resources