If anyone can query data in plain English, is the data analyst's job the SQL or knowing which question is worth asking?

The shift

Turning a question into a query and a query into a chart — SQL, joins, dashboard-building — goes from scarce, learned skill to abundant, near-instant. A model wired into the warehouse's schema, metric definitions, and query history can now translate "why did churn spike in the Nordics last month" into runnable SQL and a chart, so the gate on getting a number stops being who can write the query.

The axioms

• The person who can write the SQL and build the dashboard is the one who can answer the business's data questions, so analyst headcount and hiring screen for query and BI-tool fluency. Rests on: translating a question into working query-and-chart is scarce, learned skill.
• Self-serve analytics is gated by a data literacy the org doesn't have, so questions queue up in a central analytics team. Rests on: query-writing is a bottleneck only specialists clear.
• The value of an analyst is producing the analysis — the report, the dashboard, the cut of the data someone asked for. Rests on: producing a competent analysis is scarce work.
• Framing the actual business question — turning "are we doing well?" into a defined metric, cohort, and comparison that answers it — is the hard part of the job. Rests on: deciding what to measure and how is a scarce judgment call.
• Whether an answer is actually true — right join, right grain, no survivorship bias, a metric that means what the reader thinks it means — has to be checked by someone who understands the data and the method. Rests on: verifying correctness is scarce, and confidently-wrong output is dangerous here.
• Reading a result correctly — is this signal or noise, is the effect real, is the comparison fair — takes statistical judgment. Rests on: knowing whether a number supports a claim is scarce judgment.
• Someone is accountable when a decision is made on a number, and a bad number that drives a bad call lands on a person. Rests on: accountability requires a liable, employable human.
• Junior analysts build judgment by doing the query and dashboard grunt-work first — the reps are how they learn the data and eventually the judgment. Rests on: hands-on production is the training path to judgment.

Invalid axioms

1. The analyst is the person who can write the SQL and build the dashboard, so you hire and value query/BI fluency. Translating a stated question into a correct query and a chart is exactly the pattern-matchable, schema-grounded generation a model now does in seconds. The habit-trap: job specs, take-home tests, and leveling still screen hardest for SQL and dashboard mechanics — the part that's stopped being the differentiator — and rank the analyst who ships the most cuts of data over the one who asks the sharper question.
2. Self-serve is gated by a data literacy most of the org lacks, so questions queue at a central team. The literacy gate was the ability to phrase the query; natural-language BI moves that to anyone who can type the question. The habit-trap: teams still route routine "what's the number for X" through an analytics queue and size the team to that ticket volume, when the queue's reason to exist was the query barrier, not the questions.
3. The value the analyst delivers is producing the analysis someone asked for. When a competent first-pass analysis is near-free to generate, being the one who produced it stops being where the value sits. The habit-trap: analysts are still measured and promoted on output volume — dashboards shipped, tickets closed — the metric that made sense when producing the cut was the expensive step.

Unchanged axioms

1. Framing the actual business question is the hard, scarce part. Turning a vague "how are we doing" into the right metric, cohort, grain, and comparison — and knowing that "active users" or "retention" has three defensible definitions and which one answers this decision — is judgment with no clean pattern to match. A model will happily answer the literal question asked; deciding it's the wrong question stays human. The natural-language tool makes this more load-bearing, not less: a badly framed question now returns a confident, polished, wrong answer instantly.
2. Whether the answer is actually true still needs someone who can verify it. A model can produce a query that runs, returns plausible numbers, and charts cleanly while joining on the wrong key, using the wrong grain, silently dropping rows, or reporting a metric that doesn't mean what the reader assumes. These are precisely the errors that look right — they're the "confidently wrong" default, and here the wrongness is invisible because the output is a number, not obviously-broken text. Verification against ground truth didn't get cheaper.
3. Reading a result correctly takes statistical judgment. Is this a real effect or noise, is the sample biased, is the comparison confounded, is the trend an artifact of a definition change — a model can run the test but doesn't reliably know when the test is inappropriate or the result is being over-read. Knowing whether a number supports the claim someone wants to hang on it stays scarce.
4. Someone is accountable for the decision made on the data. A model can't own the call to reallocate budget or kill a feature on the strength of an analysis. When a decision made on a number goes wrong, a human is answerable — and that person needs to have understood the analysis well enough to stand behind it, whether or not they wrote the query.

New axioms

1. Self-serve natural-language BI produces confident, wrong answers at scale, to people with no way to catch them. The users who most need the query barrier removed are the least equipped to notice a wrong join or a misleading grain — the answer arrives polished and plausible with no error bars on its trustworthiness. The open problem is building guardrails, semantic layers, and certified-metric definitions so that the easy path returns a right answer, because "anyone can ask" now means anyone can be confidently misled.
2. Nobody wrote most of the analyses by hand, so who verifies them is unresolved. Verification used to ride along with authorship — the person who wrote the query broadly knew where it could be wrong. When analyses are generated at volume by people who didn't write them and can't read them, the review that STILL HOLDS says is essential has no natural owner, and the volume of check-worthy output outpaces the few who can actually check it.
3. The junior-analyst training rung is being pulled out from under the skill it fed. Judgment about framing, grain, and whether a result is real was built by doing the query and dashboard reps first. If that grunt-work is automated away, it's an open problem how the next analyst acquires the judgment the top of the profession still runs on — the abundant thing was the training ground for the scarce thing.

Where it breaks

"Self-serve is gated by query literacy, so route questions through a central team" (invalid) collides with "natural-language BI produces confident wrong answers to people who can't catch them" (new): orgs that dissolve or shrink the analytics team because "everyone can self-serve now" are removing the layer that caught wrong joins and misleading grains at exactly the moment they've multiplied the number of unvetted answers driving decisions — the queue is gone and so is the verification that quietly came with it.

"Value the analyst on the analysis they produce" (invalid) collides with "the junior rung that produced judgment is being automated away" (new): teams that keep leveling and hiring on production volume while letting tools do the production are cutting the reps that built the framing-and-verification judgment they still depend on — and won't feel it until the cohort that learned the old way moves on and there's no one who built the instinct to notice the number is wrong. This call hinges on how fast NL-BI tooling and semantic layers get reliable enough to shrink the verification burden; if that curve is steep, the missing training rung bites sooner than the tooling saves you.

ai-axioms.vercel.app/axioms/engineering/data-analyst-ai-axioms