Job

Can an AI agent answer your data questions and run your reports?

Researched November 2026

Can an agent do this job today?

Supervised

Agents answer business questions from company data in production at more than one company, but the evidence shows people checking the numbers. Databricks reports that Webmotors runs 26 curated Genie Agents for 353 active users, and that Mural's data team reviewed early questions to make sure its agents answered correctly. Snowflake reports that Cleverbridge uses Cortex Analyst and related tools so an agent produces client performance reports, while customer success managers interpret the results. All three accounts are vendor customer stories. The documentation of both products says answer quality depends on a curated semantic layer, and Databricks calls Genie nondeterministic. On the Spider 2.0 research benchmark of enterprise SQL workflows, the best agent framework tested solved 21.3% of tasks. Numbers that reach executives, clients or the board need a person to check them, so this job is supervised, not ready.

Abstract illustration of a row of blank ring-shaped gauges with arcs of light, the largest one glowing green

On this page

Can an AI agent do business intelligence reporting?

Yes for answering metric questions and refreshing reports, with your data team checking the numbers. AI agents for business intelligence take a plain-language question, write the SQL against your governed data, run it and return the answer. Webmotors and Mural use them in production, according to customer stories published by Databricks, and Cleverbridge does, according to Snowflake. What stays with you: defining each metric, reconciling numbers before they leave the team, and deciding what the numbers mean.

The verdict is supervised because data jobs fail silently. A wrong join or a wrong filter returns a clean-looking number. Both vendors' documentation says answer quality depends on a curated semantic layer that your team builds and maintains. On Spider 2.0, a peer-reviewed benchmark of 632 enterprise SQL workflow tasks on databases that often have more than 1,000 columns, the authors' agent framework based on o1-preview solved 21.3% of tasks, against 91.2% on the older Spider 1.0 and 73.0% on BIRD. That benchmark tests a research agent on raw schemas, not a tuned product on your data, but it shows why you check.

What the agent does and what stays with you

Agent

0of 9

Agent with review

5of 9

Answer ad-hoc metric questions asked in plain language
Write and run SQL against governed tables
Refresh recurring reports and executive packs
Drill into a KPI movement to find its drivers
Build dashboards and charts

Human

4of 9

Define metrics, joins and verified queries in the semantic layer
Interpret results and recommend actions
Grant data access and decide who can query what
Track answer accuracy and regressions over time

9 tasks: 0 agent, 5 agent with review, 4 human

Task	Who does it	Why
Answer ad-hoc metric questions asked in plain language	Agent with review	Genie and Cortex Analyst turn a plain-language question into SQL and run it, and Mural's data team reviewed early questions to confirm the answers were right.
Write and run SQL against governed tables	Agent with review	Both products generate SQL from a semantic layer, and Genie shows the generated query to users with edit rights so it can be checked.
Refresh recurring reports and executive packs	Agent with review	Webmotors says monthly executive committee reports moved from a day of manual aggregation to immediate retrieval; numbers that leave the team still need a reviewer.
Drill into a KPI movement to find its drivers	Agent with review	Webmotors product managers run root cause analysis with Genie Agents during incident reviews, and Cleverbridge uses agents to connect KPI movements to causes.
Build dashboards and charts	Agent with review	Webmotors reports that 92% of a sample of its dashboards were built with AI assistance, by business users rather than the data team.
Define metrics, joins and verified queries in the semantic layer	Human	Both vendors' documentation makes the data team responsible for this setup, and answer quality depends on it.
Interpret results and recommend actions	Human	Cortex Analyst's documentation says it does not generate insights for broad questions such as trends, and Cleverbridge keeps interpretation with its customer success managers.
Grant data access and decide who can query what	Human	Both products run queries under the user's existing permissions, which a person sets in the warehouse or catalog.
Track answer accuracy and regressions over time	Human	Snowflake's evaluations score generated SQL against verified queries automatically, but a person picks the test set and decides what to fix.

The line runs between retrieving numbers and defining them. The agent writes and runs queries. People decide what "active customer" means, which tables count, and what a change in the number implies.

Databricks reports that Webmotors, a Brazilian automotive marketplace, set up 26 official Genie Agents for business domains such as B2C revenue, churn and dealer metrics, used by 353 active users across seven departments. In August 2026, 81% of questions went to those curated agents, and negative feedback on data answers fell from 41.9% to 10.6% over 16 months. The company says monthly executive committee reports went from a full day of manual aggregation to immediate retrieval. These are the vendor's and customer's figures.

Mural's account, also published by Databricks, describes the setup cost honestly. Its Redshift warehouse held thousands of tables, many tracking the same metric with slightly different calculations. Its Director of Data Engineering said Genie "wasn't going to know which one to use" and that humans could not either. Mural rebuilt the data into one documented layer first. After launch, the data team reviewed early questions to confirm the agents answered them correctly.

Snowflake reports that Cleverbridge, a software commerce platform, got only about 60% accuracy when it scaled its own in-house pilots, then moved to Cortex Analyst and related Snowflake tools. It now has an agent produce client performance slide decks in five to ten minutes, a job that took hours, while customer success managers interpret the results. Snowflake reports over a thousand hours saved, a 33% reduction in time to pull insights together.

A worked example: one weekly revenue question

This is an illustration, not a customer story. Assume, for this example, that your sales lead asks the agent in chat: "What was net revenue in Florida last week compared with the week before?"

The agent matches "net revenue" to the approved metric in the semantic layer, which excludes refunds. It finds that the state column stores "FL", filters on it, applies your fiscal week and returns two numbers with the SQL it ran. The sales lead pastes the figure into a client email draft. Your rule says client-facing numbers get a check, so an analyst opens the SQL, sees the right metric and filter, and compares the total with the same week in the finance system. They match within rounding, and the email goes out.

For this example, assume the analyst's check takes two minutes; the visible SQL gives them a way to inspect the query. Without the value matching, Databricks' documentation shows the same kind of question producing a filter on "Florida" that returns no rows.

What you need to provide

Give the agent a read-only role on the curated tables or semantic layer it may query. Never give it write access to production data for this job. Both products run queries under existing permissions: Cortex Analyst respects Snowflake's role-based access control and needs SELECT on the tables in the semantic model, and Genie Agents follow the permissions set in Unity Catalog.

The agent touches whatever those tables hold: sales, customer and often financial records. Apply your existing row and column restrictions, and check where the vendor processes prompts and metadata. Snowflake states that Cortex Analyst does not train on customer data and, with its default models, keeps data inside Snowflake's governance boundary.

Then write down what your analysts know. That means one definition per metric with its table, filters and date logic; the fiscal calendar; and 20–50 real questions with the SQL your team would write. Databricks' documentation recommends example queries for logic unique to your organization, such as which opportunities count as pipeline. Snowflake uses the same kind of verified queries both to guide answers and to test them.

Job brief template

Use this as the brief for a vendor, an operator or your own data team. Replace the bracketed values with your own.

Copy-ready brief

Goal: Business users get correct answers to [list of recurring metric questions] from governed data within [5] minutes, and the [weekly KPI report] is refreshed and reconciled by [Monday 9:00].
Deliverable: A configured question-answering agent on the curated data, a refreshed report each period in the shared reporting folder or BI tool, and a log of every question asked with the SQL the agent ran.
Inputs: A read-only warehouse role limited to the curated tables or semantic layer the agent may query; no write access to production data
Written definitions for each metric in scope (for example revenue, active customer, churn), with the table, filters and date logic for each
20–50 example questions with the SQL your analysts would write, to use as verified queries and as the test set
The current version of each recurring report, with last period's final numbers as the reconciliation baseline
The fiscal calendar, currency rules and any business rules a new analyst would need to be told
Acceptance criteria: On a held-out set of [30] questions with known answers, at least [27] return the same result as the analyst's SQL.
Each refreshed report's headline totals reconcile to the finance or source-system figure for the same period, within [0.5%] or with the difference explained.
Every answer shows the SQL or the verified query it used, so a reviewer can trace it.
Questions outside the curated data are declined or flagged, not answered from a guessed table or column.
No query runs with write privileges; the warehouse audit log for the period shows only reads by the agent's role.
Systems: cloud data warehouse or lakehouse (for example Snowflake or Databricks)
semantic layer or metric definitions
BI or dashboard tool
team chat tool where users ask questions
Constraints: The agent reads company data that may include customer, sales and financial records; limit its role to the tables approved for this job and apply your existing row and column restrictions.
Check where the vendor processes prompts and metadata and whether it trains on your data before connecting.
Numbers sent to clients, the board or regulators go out only after human sign-off.
Volume: Example: 200–2,000 ad-hoc questions and 2–10 recurring reports a month
Cadence: Example: questions on demand; reports weekly and at month end
Human checkpoints: Metric definitions: the data owner approves each definition and verified query before users rely on it.
Access: a data platform owner grants the read-only role and reviews it when tables are added.
Reports: an analyst reconciles headline totals against the source system before each report is sent.
Accuracy: the data team reruns the held-out question set after every change to the semantic layer or model.
Conclusions: the report owner writes the commentary and recommendations.

The reconciliation criterion does the most work. A report that ties to the finance system's total for the same period is unlikely to have a wrong join hiding in it. A report that does not tie tells you before a client does.

Risks and human checkpoints

The agent picks the wrong table or join and returns a plausible but wrong number.
How you notice it
Show the generated SQL with every answer and reconcile headline totals to a known figure; Mural rebuilt thousands of duplicate tables into one documented layer before opening its agents to staff.
A user's wording does not match the stored values, so a filter silently returns nothing or the wrong rows.
How you notice it
Test questions that use everyday names for codes; Databricks' documentation shows an agent filtering on 'Florida' when the data stores 'FL' until value matching is turned on.
Two definitions of the same metric give two answers, and the agent's answer changes from day to day.
How you notice it
Keep one approved definition per metric in the semantic layer and track repeat questions; Databricks says Genie is nondeterministic and conflicting guidance raises the risk of undesirable responses.
A model or configuration change makes previously correct answers wrong.
How you notice it
Rerun the held-out question set after each change and watch the regression count; Snowflake notes that model selection may change and different models produce different results.
The agent is asked a broad question it cannot answer from SQL and returns something that reads like an insight.
How you notice it
Sample answers to open questions weekly; Cortex Analyst's documentation says it does not generate insights for broad questions such as trends.
The agent's role can see or change more data than the job needs.
How you notice it
Review the role's grants before launch and audit query logs for writes or unexpected tables.

Human checkpoints

Metric definitions: the data owner approves each definition and verified query before users rely on it.
Access: a data platform owner grants the read-only role and reviews it when tables are added.
Reports: an analyst reconciles headline totals against the source system before each report is sent.
Accuracy: the data team reruns the held-out question set after every change to the semantic layer or model.
Conclusions: the report owner writes the commentary and recommendations.

The common thread is that wrong answers look like right ones. A chart built on the wrong table renders as cleanly as a chart built on the right one. Visible SQL, one definition per metric and a standing test set are what turn silent failures into visible ones.

Snowflake's documentation lists limits worth knowing before rollout: Cortex Analyst cannot use the results of a previous query in a follow-up, struggles with long conversations that shift intent, and answers only questions that SQL can resolve.

How to run a trial

Test the agent on questions you already know the answers to.

Collect 30 questions your analysts answered last quarter, with the SQL and the result for each.
Configure the agent on the curated data with your metric definitions and a separate set of example queries. Keep the 30 test questions out of its examples; Snowflake's evaluation feature removes test queries from the semantic view for the same reason.
Ask the 30 questions and compare results. Count wrong numbers and wrong refusals separately.
Refresh last period's recurring report and reconcile the headline totals with the final version.
Ask five questions outside the curated data. The agent should decline or flag them.
Change one definition, rerun the test set and count regressions.

Collect 30 questions
Your analysts answered them last quarter, with the SQL and the result for each
Configure the agent
On the curated data with your metric definitions and separate example queries, keeping the 30 test questions out
Ask the 30 questions and compare results
Count wrong numbers and wrong refusals separately
Refresh last period's recurring report
Reconcile the headline totals with the final version
Ask five questions outside the curated data
The agent should decline or flag them
Change one definition
Rerun the test set and count regressions

If it passes, open it to one team for a month with the data team reviewing a sample of answers each week, as Mural did at launch. Widen access when the error rate stays low.

On cost, Snowflake bills Cortex Analyst per successful message at a credit rate in its consumption table, plus warehouse costs for running the generated SQL. Databricks says Genie One and Genie Agents usage by users is free through January 31, 2027, while service principals are billed. Ask either vendor what your expected question volume would cost.

For the brief, see how to write a job description for an AI agent, and for structuring the test, how to run a paid trial with an AI agent. Related pages: data engineering, spreadsheet analysis, the data analyst role and the data overview.

Sources

Genie
Databricks documentationPrimaryAccessed Nov 14, 2026
Tune Genie Agent quality
Databricks documentationPrimaryAccessed Nov 14, 2026
Cortex Analyst
Snowflake documentationPrimaryAccessed Nov 14, 2026
Cortex Analyst evaluations
Snowflake documentationPrimaryAccessed Nov 14, 2026
Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows
ICLR 2025 proceedings (Lei et al.)PrimaryAccessed Nov 14, 2026
Webmotors accelerates data queries for 353 users with Genie Agents on Databricks
DatabricksVendorAccessed Nov 14, 2026
Mural answers data questions in minutes with Genie Agents
DatabricksVendorAccessed Nov 14, 2026
Cleverbridge Uses Snowflake Cortex AI to Accelerate Insights Across $1B+ in Global Software Commerce
SnowflakeVendorAccessed Nov 14, 2026

What we could not verify

All production evidence on this page comes from customer stories published by the two platform vendors (Databricks and Snowflake). No independent audit of answer accuracy in a company's production BI setup was found.
Webmotors' time savings come from an internal user survey and audit log estimates reported in a Databricks customer story; the method is not published.
No public accuracy figure was found for either product on a customer's own data; Spider 2.0 measures research agents on a benchmark, not these products.
Agent pricing is not shown as a dollar range: Snowflake bills Cortex Analyst per message in credits through its consumption table, and Databricks lists Genie One and Genie Agents as free for users through January 31, 2027. No human wage benchmark was researched, so the page has no cost comparison.

Use this brief

The AI agent job board isn't open yet. Tell us about the job and Nikolas will reply personally.

I build or operate AI agents