By James Farrelly | May 2026

A colleague reached out a few weeks ago with a question I have been thinking about ever since. His 18-year-old son is heading to college in September. He wants to study data science and AI. The father’s instinct was that this sounds like a good choice, but he had a nagging doubt: will AI do data science better than his son ever could?

I told him I did not have a clean answer. And I meant it. But the more I have sat with it, the more I have come to believe that the nagging doubt is pointed at the wrong thing.

The work is not disappearing. It is shifting.

Dan Shipper, the co-founder of Every.to, published a piece recently called After Automation that articulated something I have been watching happen inside Pfizer for the last two years. His team at Every has automated everything they can – agents write code, draft emails, handle support, compile newsletters. They go, in his words, “as far and as fast as possible.”

Yet there is more human work to do than ever.

This matches my own experience almost exactly. Every process, every department, every team I work with is actively asking what AI can do to improve the flow of work. In our case, the goal is to speed the scientific process of developing breakthrough medicine and getting it to patients faster. The urgency is real. The investment is real. The output is increasing.

But so is the work.

AI is a force multiplier, which means the humans utilizing it need to know what they are multiplying. Shipper’s framing is useful here: technological progress creates more for people to do, not less. When cheap competence floods in, it creates sameness. The human value shifts to the frame, not the execution.

What Lenny saw happening in the industry

In May 2026, Lenny Rachitsky, one of the most widely-read voices in product and technology, posted a thread on X that sparked a few hundred thousand impressions and a wave of conflicting responses. His observation was direct: AI has democratized data analysis to the point where product managers and engineers can now generate their own reports instantly, using LLMs. The data scientist’s daily job, in many organizations, has shifted. Instead of proactive analysis and modeling, they are increasingly acting as auditors, spending their time cleaning up and debunking AI-generated metrics that are wrong roughly half the time.

The responses broke into three camps:

The first camp was optimistic. AI handles the low-stakes questions so data scientists can do harder work. The backlog gets shorter. The SQL monkey-work disappears. That is a good thing.

The second camp was frustrated. Product managers are using AI to find a chart that proves their point, without understanding that the underlying query joined the wrong tables or that the metric was hallucinated entirely. Data scientists have become the “bad guys” who continually prove why everyone else’s fast-moving numbers are inaccurate. That is exhausting.

The third camp saw a structural shift. The traditional analyst role – pulling data, building dashboards – is being absorbed by AI-augmented business roles. The value of a data scientist is moving to infrastructure, data governance, and the deep expertise required to know what to measure and how to measure it correctly.

All three camps are right about something.

The trust problem no one is solving fast enough

I keep coming back to this in my own work. The systems I am building can do a meaningful level of data science. They can reason across complex datasets. They produce outputs. They produce numbers.

The problem is that a number on a page is not trusted the way an experienced data scientist is trusted.

When a data scientist tells me that a metric moved, I know that person has lived inside the problem. I know they have thought about the data pipeline, the assumptions in the model, the edge cases that break the query. I know they are accountable for the interpretation, not just the output.

When an AI agent produces a number, the instinct is to ask: where did this come from? What transformation was applied? What assumptions were made? Was the right table joined? Was the right time window used?

Some of those questions, AI can now answer. The provenance of a calculation, the chain of reasoning, the intermediate steps – these are increasingly traceable if the system is designed well. That is real progress, and it matters.

But many of the harder problems do not get a clean answer from the machine. They get a quick answer that prompts deeper analysis. Some problems are just genuinely hard to solve, and the speed of an AI-generated first pass can create a false confidence that you are already close to the answer when you are not.

This is where the new work for data scientists lives. Not just producing analysis. Building trust in systems. Evaluating agentic processes. Designing the guardrails that make AI-generated numbers trustworthy to the people who have to act on them.

Can I prove where this number came from? Can I show what assumptions were encoded in the system that produced it? Can I demonstrate that the output is meaningful and not an artifact of a broken query or a hallucinated metric?

Those are data science questions. They are not getting easier.

What I would tell my colleague’s son

For me, it was computer science. I loved it. I built a career around it. That field has changed more in the last three years than in the previous three decades, but the love of it is still what gets me in the game.

Choosing data science as a degree means you are curious about how numbers tell stories. That curiosity is not obsolete. The people I work with who are doing this well – who are building the pipelines, auditing the outputs, designing the governance structures, and earning the trust of the teams that depend on their analysis – are not going to be replaced. They are going to be amplified.

AI will empower the next generation of data scientists to learn faster, do more, and move with a scale that was not available to anyone before. But the judgment, the domain knowledge, the ability to look at an output and ask the right skeptical question – that is still a human job.

Data science is not going away. It is just getting harder to do well. Which, if you love it, is actually the best possible news.

---

James Farrelly is Head of Agent Product Management at Pfizer’s Commercial AI Acceleration team. He writes at james-farrelly.com.