Loading...
The honest answer from inside the job: AI did not shrink the work, it multiplied the things that generate work. Play reviewer for three robot-written queries and see why.
Conor Robertson
August 16, 2026 · 7 min read
The questions people actually type into Google at 11pm.
QueryCase teaches SQL through mystery cases on real datasets. Music, sports, film, and more. Your first cases are completely free.
Start your first caseFree to start. No credit card required.
There has never been a better time to ask this question, because half the answers you will find are written to scare you and the other half are written to sell you something. I run a platform that teaches SQL, so read me with that in mind too. But I watch what companies actually do with data every week, and my honest answer is this:
No, AI will not replace data analysts. It has already done something more interesting: it changed what the job is.
Writing syntax is no longer the skill. An AI will draft you a query in two seconds, and most of the time it will run. If your entire value was turning a clearly written ticket into a clearly specified query, that translation layer is now automated, and pretending otherwise helps nobody.
That is the part every doom article stops at. It is the least interesting part.
Here is what the replacement narrative misses. AI did not just speed up analysts. It sped up everyone who creates the things analysts measure.
Engineers ship more code than they did two years ago. Product teams spin up more features, run more experiments, change more flows. Every one of those changes lands with the same questions attached: did it work, who did it affect, what moved, was it real or noise.
Demand for analysis scales with the pace of change, and AI just permanently increased the pace of change. That is why the same companies adopting AI everywhere are still hiring people who can read data properly. The queue of questions is growing faster than the tools that answer them.
The idea that stakeholders will answer their own questions has been ten years away for about fifteen years. AI chat-with-your-data tools are the newest version of the promise, and they hit the same wall the dashboards did: most companies are nowhere near ready for self-serve analytics.
Real warehouses are messy. Tables are duplicated, half-documented and quietly wrong. Two revenue columns disagree and only one person knows why. Point a chatbot at that and it does not say "your data is ambiguous". It picks one interpretation and answers confidently. A stakeholder cannot tell the confident right answer from the confident wrong one, which is precisely why they asked someone else in the first place.
Somebody has to own correctness. That somebody is the analyst.
This is the change I find genuinely hopeful, and it is the one nobody puts in a headline: clean schemas and semantic layers matter more than they ever have.
AI is only as accurate as the structure underneath it. Against a well-modelled warehouse, a star schema with honest definitions and a semantic layer that says what "revenue" means, AI-generated SQL is often right, and so is everyone else's. Against a swamp, it is a confident nonsense machine. The same is true of humans, but humans at least get suspicious.
So the skill that appreciates fastest is modelling: knowing what a fact table is, why grain matters, how to define a metric once so a thousand questions inherit the definition. The analysts who can do that are not competing with AI. They are building the thing that makes AI usable.
Generated queries fail in ways that do not look like failure. The query runs. The number appears. The number is wrong. Here are three queries the robot wrote against tiny datasets, small enough that you can check every row yourself. Two contain a real bug. Would you have caught them?
That is the quiet irony of AI writing SQL: it makes reading SQL more valuable, not less. The only defence against a plausible wrong answer is a person who reads the query and catches it, the way an editor catches a plausible sentence that happens to be false. Companies know this, which is why they still put candidates in front of a live SQL exercise. It is worth practicing under those exact conditions: the interview is not testing whether you can produce syntax, it is testing whether you can be trusted to sign off on an answer.
Honestly, the work got better. The hours that used to disappear into boilerplate now go into the parts that were always the point: working out what question actually matters, structuring data so answers are cheap, reading an experiment honestly, and saying "that number is wrong and here is why" in a room where nobody else can tell.
The short version
Less typing, more judgment. The analyst of 2026 frames the question, owns the model, supervises the machine and vouches for the number. AI removed the boring third of the job and raised the stakes on the rest.
If you are choosing a direction, the three skills that compound are SQL fluency deep enough to verify anything put in front of you, data modelling, and experiment literacy. If you are still mapping where you fit, the data roles guide shows where each role sits and the SQL each one actually writes.
The people with a reason to worry are the ones who only ever translated tickets into queries. The people with a reason to be optimistic are the ones learning to be the judgment in the loop. That skill is learnable, it is practicable, and it is exactly the one that gets tested when it counts.