Am I going to lose my data job to AI?

Almost every week I speak to a data analyst or a data leader who asks me, sometimes directly and sometimes between the lines, whether AI is coming for their job. It's a fair question. Their feed is full of people announcing the death of the analyst, usually written by someone selling something. Probably mine included.

So I did what a data person would do and looked at what actually happened the last few times a machine learned to do something humans were paid for. The story turns out to be more interesting than both the doomers and the hype merchants make it sound.

The spreadsheet was supposed to kill the accountant

In 1979 a piece of software called VisiCalc came out. It could recalculate an entire ledger in seconds, work that used to take a person a full day with a pencil and a large eraser. If any technology was going to wipe out a profession, this was it, because the arithmetic was most of what clients could see accountants doing.

Since 1980, around 400,000 bookkeeping and accounting clerk jobs disappeared in the US. Over the same period, around 600,000 accountant jobs were added. The arithmetic went away and the accountants multiplied.

Why? Because once the calculation was free, clients started asking for more of it. Suddenly you could play what-if. What if we made the product 2% bigger? What if we gave everyone a raise? What if we borrowed more? Accounting got cheap enough to use for everything, so people used it for everything. The people who were purely doing the arithmetic lost out, and the people responsible for what the numbers meant became more valuable than they'd ever been.

The ATM was supposed to kill the bank teller

Economist James Bessen studied what happened to bank tellers when ATMs rolled out across the US. Everyone, including bank managers, assumed the teller was finished. A machine that dispenses cash and takes deposits is the job, right?

The ATM cut the number of tellers needed per branch from about 21 to 13. But that made branches much cheaper to run, so banks opened 43% more of them. Fewer tellers per branch, more branches, and total teller employment actually grew for decades while 400,000 ATMs were being installed. The job changed underneath the title: less counting cash, more handling exceptions, relationships and judgement, the things the machine was hopeless at.

Mind you, teller numbers did eventually fall, years later, when mobile banking removed the reason to visit a branch at all. So I'm not claiming automation never touches jobs. What the data shows is that automation kills tasks immediately and jobs slowly, and that the people who moved toward the judgement side of the work did fine, while the ones who kept defining themselves by the automated task didn't.

The AI was supposed to kill the radiologist

This one is my favourite because it's the closest to what data people are being told right now. In 2016, Geoffrey Hinton, the godfather of deep learning himself, stood on stage and said people should stop training radiologists immediately, because it was "completely obvious" AI would outperform them within five years. He compared them to the coyote that's already run off the cliff and hasn't looked down.

A decade later, the US has a historic radiologist shortage. Thousands of open roles taking months to fill, salaries at record highs, and the Mayo Clinic's radiology department has grown by more than half since the prediction. Hinton has since walked it back, clarifying he only meant the image-reading part.

That clarification is exactly the point because reading the scan was never the whole job. The job is deciding what the finding means for this patient, consulting with the physicians, guiding treatment, carrying the liability for being wrong. AI genuinely got good at the visible task, and the profession got bigger anyway, because reading scans was never what made a radiologist valuable.

Sound familiar?

I hear a version of this in investor meetings all the time. Some flavour of "Claude can write SQL and Python now, so what do data teams even do? And for that matter, why does Querio need to exist?" I understand why they ask. From where they sit, a data analyst answers questions with data, an LLM can write SQL, and so both the profession and my company look redundant. It's the same mistake Hinton made, just with a term sheet attached: describing the job from too far away.

Up close, the job is knowing that revenue means three different things depending on which team is asking. It's knowing the events table double-counts before March because of a pipeline migration nobody documented. It's knowing that when the CEO says "the Q2 launch" they mean the thing that actually shipped in July. It's telling someone the number they asked for is technically correct and practically misleading.

None of that is in the schema. Most of it isn't written down anywhere. It lives in people's heads, and it's exactly what makes an answer accurate rather than just confident. An LLM with none of that context doesn't replace an analyst. It replaces a very fast intern who has never worked at your company, and who will confidently give you wrong answers that take weeks to surface.

Writing SQL is the data person's version of the arithmetic, the cash-counting and the scan-reading. It's the visible part of the work but it was never the valuable part.

So what actually changes?

The nature of the work changes not the purpose. Your purpose has always been making sure the company has accurate answers, analytics people trust, reports that hold up, and data science that's actually right. That was the job before SQL existed and it'll be the job long after "prompting" stops being a word anyone uses.

What changes is where you spend your time. Anthropic recently published a detailed writeup of how they enabled self-service analytics internally, and it's worth reading closely, because the most important numbers in it aren't the headline ones. Out of the box, pointed at their warehouse, Claude answered analytics questions correctly about 21% of the time. To get above 95% they needed a cleaned-up dimensional model, a semantic layer with human-owned metric definitions, curated documentation of tables, joins and gotchas, and evaluations wired into CI. They even tried having the model auto-generate the metric definitions from raw tables and query logs, and it made accuracy worse, because it faithfully encoded the exact ambiguities a human would have caught. And when they briefly relaxed the maintenance, accuracy fell from roughly 95% to 65% in a month.

Think about what that actually means. The frontier AI lab, with unlimited access to its own frontier model, got from 21% to 95% by having humans encode their judgement into governed, maintained artefacts, and the system started rotting the moment those humans stopped. To me that doesn't read like a profession being replaced. It reads like the clearest job description for the next decade of data work that anyone has published. Who do you think maintains all of that? That's my answer to the investors, too.

The analyst's job shifts from writing the thousandth query to encoding the knowledge that makes ten thousand queries correct, from answering questions to owning the system of answers. And honestly, this is the promotion the profession has been asking for. Nobody became an analyst because they loved being a human SQL vending machine for the rest of the company. They became analysts because they liked finding out what's true and helping people act on it. The vending machine part is what's going away.

And like the spreadsheet did for accountants, cheap answers will explode demand. When any question costs seconds instead of a ticket and a three-day wait, people ask far more questions. We see this in our own numbers at Querio: after teams implement it, they end up asking roughly three times more questions of their data than before. Those questions didn't come from nowhere. They always existed, they just died in the queue, or in someone's head, because asking wasn't worth the wait. The demand for answers was never the constraint, the cost of answering was. And someone has to make sure three times more answers are right.

My honest opinions

1. If your entire job is translating English into SQL with no context, judgement or ownership attached, you're the bookkeeper in 1980. Be honest with yourself and move up the stack now, while it's still a choice.

2. Business context is getting more valuable, not less. When machines can generate numbers instantly, the person who knows why the numbers are the way they are becomes the constraint, in a good way.

3. Trust becomes the product. When answers get cheap, trusted answers get precious. Your name on a number will mean more, not less.

4. The teams that shrink will be the ones that defined themselves by output volume. The teams that grow will be the ones that defined themselves by accuracy and ownership.

5. Nobody will hand you time to make this transition. Take it. Automate your own drudgery before someone else automates it and calls it a restructure.

One last thing

Getting rid of the data team won’t get rid of the work. The questions still get asked and the numbers still get generated, except now nobody is checking whether they're right. Companies will learn this the way companies always learn things: in a board meeting, from a number that doesn't survive a second look.

The spreadsheet didn't reward the companies that fired their accountants, it rewarded the ones that started asking them better questions. I think the exact same thing is about to happen to data teams. Your job is not going anywhere. It's finally becoming the job it was always supposed to be.

If you're a data person going through this shift and want to compare notes, or you just disagree with me, email me directly: rami@querio.ai.