# Data analyst

> Turns a business question into a query, a query into a number, and a number into a decision someone will act on.

Page: https://flyvolo.ai/en/careers/data-analyst
Impact index: 63/100 · confidence Medium confidence · assessed 2026-09-10
This is not a probability of losing a job.

## Scope

Written for business, product and marketing analysts who work in SQL, spreadsheets and BI tools. Data engineers, data scientists building models, and statisticians in research settings differ.

## Tasks

### Writing queries and building dashboards

- Direction: Automating
- Basis: evidenced (verified records attached)
- Why: Natural-language-to-SQL and chart generation are among the most mature applications of language models, and the BI vendors have built them into the products analysts already use. The mechanical translation step that filled much of an analyst's week is now a prompt, when the data is clean and the question is clear.
- What this does NOT establish: 'When the data is clean and the question is clear' is a large condition. In most companies neither holds, and the tools produce confident wrong answers on ambiguous schemas.
- Records:
  - 2026-09-10 · Capability · OpenAI published a data-agent template whose own instructions require the metric definitions, trusted queries and corrections to be supplied by the team using it — https://flyvolo.ai/en/changes/ev-20260910-data-analyst-3 (source: https://developers.openai.com/cookbook/examples/agents_api/apps/data_analyst/readme)
  - 2026-07-20 · Constraint · On Beaver, a text-to-SQL benchmark built from real corporate warehouse query logs, a plain LLM scored zero and an agentic setup reached about 10%, against 80-90%+ on the public benchmarks — https://flyvolo.ai/en/changes/ev-20260720-data-analyst-2 (source: https://cacm.acm.org/blogcacm/if-you-think-you-can-do-real-world-text-to-sql)
  - 2024-09-19 · Deployment · Uber put QueryGPT, a natural-language-to-SQL tool, into production for operations and support teams, reporting query authoring time down from about 10 to about 3 minutes (~300 daily users) — https://flyvolo.ai/en/changes/ev-20240919-data-analyst-1 (source: https://www.uber.com/en-US/blog/query-gpt/)

### Working out what is really being asked

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: The hard part of analysis is upstream of the query: knowing the organisation's definitions, its politics, and what decision the number will feed. A tool answers the question as asked; the analyst's value was always in noticing the question was wrong.
- What this does NOT establish: Self-service tools let managers ask their own questions badly rather than ask an analyst. Framing stays valuable, but the analyst has to be in the room to supply it, and fewer are.
- Records:
  - 2025-07-18 · Deployment · LinkedIn says its text-to-SQL bot has over 300 weekly users, that owners' annotations lift correct or nearly correct answers from 9% to 48%, and wrong filters fail most — https://flyvolo.ai/en/changes/ev-20250718-data-analyst-4 (source: https://arxiv.org/abs/2507.14372)

### Knowing when the data is lying

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: This requires history with the specific data — remembering what happened to it and why — which is exactly what a model without organisational memory lacks. Faster query generation makes this more important, not less, because wrong answers now arrive faster and look more polished.
- What this does NOT establish: Institutional memory of the data is built by working with it daily. If the daily querying moves to self-service, the memory that makes this task possible stops accumulating.
- Records:
  - 2026-09-10 · Capability · OpenAI published a data-agent template whose own instructions require the metric definitions, trusted queries and corrections to be supplied by the team using it — https://flyvolo.ai/en/changes/ev-20260910-data-analyst-3 (source: https://developers.openai.com/cookbook/examples/agents_api/apps/data_analyst/readme)
  - 2026-07-20 · Constraint · On Beaver, a text-to-SQL benchmark built from real corporate warehouse query logs, a plain LLM scored zero and an agentic setup reached about 10%, against 80-90%+ on the public benchmarks — https://flyvolo.ai/en/changes/ev-20260720-data-analyst-2 (source: https://cacm.acm.org/blogcacm/if-you-think-you-can-do-real-world-text-to-sql)

### Making a number change a decision

- Direction: Being augmented
- Basis: evidenced (verified records attached)
- Why: Drafting the narrative and the slides is faster. Persuading a sceptical director in a meeting, and knowing which caveat to lead with for this particular audience, remains a human task because it depends on reading the room and on the analyst's own credibility.
- What this does NOT establish: Credibility is personal and accrues slowly, which protects established analysts and no one else. A first analyst hire has none of it and now has fewer routine tasks to earn it with.
- Records:
  - 2026-06-25 · Constraint · The UK statistics regulator says producers using AI to produce or communicate official statistics should apply the Code of Practice as for any method, with human oversight and clear accountability — https://flyvolo.ai/en/changes/ev-20260625-data-analyst-6 (source: https://osr.statisticsauthority.gov.uk/publications/state-of-the-statistical-system-2026-2/pages/6/)

### Owning the definitions the tools rely on

- Direction: New task
- Basis: evidenced (verified records attached)
- Why: Natural-language analytics only works on top of well-defined metrics, so the value of the analyst shifts from answering questions to curating the layer that lets everyone answer their own. This is new work, closer to data governance than to reporting, and it is where analyst headcount is being redirected.
- What this does NOT establish: Curating definitions is a smaller job than answering questions was, and it is one per organisation. Headcount being 'redirected' into it is a euphemism for a smaller team doing different work.
- Records:
  - 2026-09-10 · Capability · OpenAI published a data-agent template whose own instructions require the metric definitions, trusted queries and corrections to be supplied by the team using it — https://flyvolo.ai/en/changes/ev-20260910-data-analyst-3 (source: https://developers.openai.com/cookbook/examples/agents_api/apps/data_analyst/readme)
  - 2025-07-18 · Deployment · LinkedIn says its text-to-SQL bot has over 300 weekly users, that owners' annotations lift correct or nearly correct answers from 9% to 48%, and wrong filters fail most — https://flyvolo.ai/en/changes/ev-20250718-data-analyst-4 (source: https://arxiv.org/abs/2507.14372)
  - 2024-09-26 · Deployment · Grab says AI-drafted documentation, sent to the tables' owners to accept or edit, raised documentation of its most-queried tables by 70 points to about 90% before it launched an LLM data-discovery bot — https://flyvolo.ai/en/changes/ev-20240926-data-analyst-5 (source: https://engineering.grab.com/hubble-data-discovery)

## How to cite this

- The unit is the task, not the job title: say which tasks are changing and which are not.
- Carry each task's basis and its limits with it; an inferred judgement is not a finding.
- This page gives no number of years. Do not derive one from it.
- More: https://flyvolo.ai/llms.txt