Can an AI agent do a data analyst's job? Task by task

Short answer

Can AI replace data analysts? Not as a whole job, based on the public evidence as of November 2026. The products cited here document AI doing the hands-on half of the work: cleaning data, writing SQL and Python, running statistical analysis and drawing charts. In each case the documentation either tells you to review the output or relies on definitions a data team sets up first. The other half of the job, deciding which question matters, building and validating models, and turning results into a recommendation someone acts on, has no cited evidence of an agent doing it.

So if you manage analysts, the realistic change is that routine questions can be answered by a tool sitting on governed data, and your analysts spend more of their time defining metrics, checking answers and working with the people who asked. Whether that means you need fewer analysts is not something the sources establish.

Some public data for scale. O*NET has no occupation titled data analyst; its closest listings are data scientists (15-2051.00) and business intelligence analysts (15-2051.01), and O*NET reports wage and employment data for both under data scientists. The U.S. Bureau of Labor Statistics (BLS) puts the national median annual wage for data scientists at $120,230 for May 2025; O*NET, citing BLS, gives $57.80 an hour and 275,600 employees. BLS projects employment of data scientists to grow 35 percent from 2025 to 2035, with about 24,800 openings a year on average. Those are figures for the occupation, not for any one team.

Task by task

The table below starts from the 17 task statements O*NET lists for data scientists, merged into 10 tasks an analytics manager would recognize. A task is marked for an agent only where a cited product documents an agent doing it. Where O*NET lists the task but no cited source shows an agent doing it, the task is marked human, and the note says so.

Agent

0of 10

Agent with review

5of 10

• Clean and prepare raw data for analysis
• Query and analyze data with SQL or statistical software
• Create charts and dashboards to show results
• Spot trends, relationships and outliers in the data
• Write code and functions to run analyses

Human

5of 10

• Build, test and compare predictive models
• Identify business problems and interview stakeholders
• Present findings and recommend data-driven decisions
• Design surveys and sampling plans
• Keep up with new analytic methods and research

On that standard, 0 of 10 tasks are ones an agent completes with only spot checks, 5 of 10 are ones a documented tool does with a person reviewing the output, and 5 of 10 stay with a person. The five reviewable tasks are preparing data, querying and analyzing it, charting it, spotting trends and outliers, and writing analysis code.

The evidence comes from the products below.

• ChatGPT data analysis. OpenAI's help center says ChatGPT can summarize "rows, columns, trends, and outliers", create tables and charts, and run "Python-based calculations, transformations, and statistical analysis" on uploaded files. It tells you to "review the generated code, outputs, and assumptions before relying on the result," and says ChatGPT may choose a method or chart type that does not match your intent on the first try.
• Snowflake Cortex Analyst. Snowflake documents it as a managed feature that lets business users ask questions in natural language and get answers without writing SQL. It relies on a semantic model or semantic view in which your team defines metrics, relationships and join paths, and runs queries under Snowflake's role-based access controls. Snowflake lists limits: it cannot use the results of earlier queries in a conversation, and it "does not generate insights for broader business-related queries, such as 'What trends do you observe?'"
• Databricks Genie. Databricks documents Genie One as a place for business users to ask data questions in natural language, Genie Agents as environments where data teams configure "trusted data, metrics, and business rules," and Genie Code as an AI coding and data assistant.

Is it used in production? In a Databricks customer story, Webmotors, which Databricks describes as a Brazilian automotive platform, reports 353 active users asking questions across 26 Genie Agents curated by its Insights team. In an internal survey, 85% of users reported saving analyst hours, averaging 3.5 hours a week, and the company estimates from audit logs that about 504 hours a month return to the business. These are the vendor's and customer's figures; the method is not published.

How reliable is it on hard problems? Spider 2.0, a benchmark published in the ICLR 2025 proceedings, built 632 text-to-SQL workflow problems from enterprise databases, often with over 1,000 columns. Its authors report that their code agent framework based on o1-preview solved 21.3% of the tasks, against 91.2% on the older Spider 1.0 benchmark. That is a research benchmark, not a test of the products above, but it shows why a query on an unfamiliar, messy schema needs checking.

A worked example

This is an illustration, not a reported case. Assume, for this example, that a sales director asks, "What was net revenue by region last quarter?" A tool on your warehouse writes the SQL from a semantic model, returns a table and a bar chart. The analyst compares the total with the finance team's closed-quarter figure and finds it is 4% higher. The query summed gross order value because "revenue" in the semantic model was never defined as net of refunds. The analyst fixes the metric definition, reruns the question, and the total now matches finance. The tool did the query and the chart. The person caught a wrong definition that would have repeated in every later answer.

Jobs you can hand to an agent today

The reviewable tasks above group into separate jobs, each with its own page covering what to give the agent, a brief you can copy, risks and how to run a trial.

• JobSupervised
  AI agents for business intelligence and reporting

  An AI agent can answer plain-language metric questions, write the SQL and refresh recurring reports from your governed data. Your data team still defines the metrics, checks numbers that leave the company and owns the conclusions.
• JobAssist only
  AI agent for spreadsheet analysis: what it can do

  An AI agent inside Excel can summarize a dataset, add formula columns and PivotTables, flag outliers and trace formula errors. Both vendors cited here tell you to verify the output, so a person checks the numbers before anyone relies on them.

Ready

0of 2

Supervised

1of 2

• AI agents for business intelligence and reporting

Assist only

1of 2

• AI agent for spreadsheet analysis: what it can do

Not yet

0of 2

• Reporting and BI covers answering metric questions and refreshing recurring reports from governed data.
• Spreadsheet analysis covers cleaning, analyzing and charting data from a file.

If you are deciding where to start, pick a recurring report whose numbers someone already reconciles, so a wrong answer shows up against a known total. We recommend giving the tool a read-only role, writing down the definitions of the five or ten metrics people ask about most before the trial, and counting how many answers match the reconciled figure on the first few weeks of questions.

What stays human

Four groups of work stay with a person in the sources read for this page.

Defining metrics and checking answers. Both platform tools cited above depend on definitions a data team writes: Snowflake's semantic views hold metrics, relationships and join paths, and Databricks' Genie Agents hold the "trusted data, metrics, and business rules" business users rely on. Webmotors' own analyst is quoted in the customer story saying Genie works well because the data is cataloged "with well-written table descriptions and standardized taxonomy." OpenAI tells you to review code and assumptions. Someone who knows the business has to own those definitions and test the answers against known numbers.

Framing the question. Identifying business problems that data can address and interviewing stakeholders are O*NET tasks with no cited agent evidence. Snowflake's own limits, no insights for broad questions such as what trends you observe, point the same way: the tool answers a well-posed question, and someone has to pose it.

Modeling. Testing and validating predictive models, applying feature selection and comparing models on performance metrics are O*NET tasks with no cited production evidence of an agent doing them.

Recommending and presenting. Presenting results to management, recommending data-driven decisions, and designing surveys and samples are O*NET tasks with no cited agent evidence. We recommend that the analyst who checked the numbers is the one who presents them.

A setup the cited documentation supports is AI tools answering routine questions and drafting analysis on governed data, and analysts defining the metrics, checking the answers and doing the work that starts and ends with people. If you have a reporting job you want an agent to take on, send us your brief below. The guide on how to hire an AI agent covers scope, trials and acceptance.

Sources

1. 1
   15-2051.00 - Data Scientists

   O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026

2. 2
   15-2051.01 - Business Intelligence Analysts

   O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026

3. 3
   OEWS national annual median wage, Data Scientists (series OEUN000000000000015205113)

   U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026

4. 4
   Data Scientists, Occupational Outlook Handbook

   U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026

5. 5
   Data analysis with ChatGPT

   OpenAI Help CenterPrimaryAccessed Nov 14, 2026

6. 6
   Cortex Analyst

   Snowflake documentationPrimaryAccessed Nov 14, 2026

7. 7
   Genie

   Databricks documentationPrimaryAccessed Nov 14, 2026

8. 8
   Webmotors accelerates data queries for 353 users with Genie Agents on Databricks

   DatabricksVendorAccessed Nov 14, 2026

9. 9
   Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows

   ICLR 2025 proceedings (Lei et al.)PrimaryAccessed Nov 14, 2026