Companies face many daily questions:
Why did sales drop this month?
Which region suddenly saw a rise in cancellation rates?
We spent more on ads, so why didn't orders rise accordingly?
Which customer segments are most likely to renew?
Customer service requests surged: which product is causing issues?
These questions usually aren't due to a lack of data.
The challenge lies in data scattered across:
Data Warehouses.
CRMs.
Spreadsheets.
Dashboards.
Google Drive.
SharePoint.
Sometimes even across different departments.
The real difficulty is:
Someone has to find, query, align definitions, analyze, and organize the data into comprehensible results for everyone.
OpenAI’s newest offering:
Data agent
aims to embed this workflow directly into ChatGPT Work.
It’s not just "dumping Excel into ChatGPT"
Data agent connects to officially approved enterprise data sources.
Currently supported sources include:
Amazon Redshift.
Google BigQuery.
ClickHouse.
Databricks.
MongoDB.
Snowflake.
And others.
On the document side, it can also bring in data from Google Drive and SharePoint for analysis.
This means usage no longer relies on:
Downloading CSV files.
Uploading them.
Then asking AI.
Instead,
AI works directly within the company’s authorized data environment.
The simplest use is just asking questions
For example:
"Why did Weekly Active Users drop last week?"
Before, you’d probably:
Ask a Data Analyst.
The analyst would confirm:
Which product?
Which market?
Compared to which period?
Then they’d write SQL, run queries, organize results, create charts, and finally explain the findings to you.
Data agent lets users start directly from the:
Business Question.
The AI then fetches data, compares periods, segments customers, identifies drivers, and organizes evidence back to you.
The key is continuous follow-up, not just a single chart
For example, if AI spots retention dropped this month, you can continue asking:
Which market fell the most?
Is it new or returning customers?
Which product versions show changes?
Did support tickets increase simultaneously?
Did similar patterns appear last month?
You don’t need to make a new report request for each question.
All investigations stay within the same conversation.
It can turn into interactive dashboards
After analysis, Data agent can create:
Interactive dashboards—not just static charts.
Teams can edit, share, refresh, and refine them.
For instance, an initial dashboard showing overall company revenue can be broken down by:
Region.
Product.
Customer segment.
Acquisition channel.
The same analysis continues to drill down deeper.
No need to abandon existing BI tools
This is practical.
Many companies have spent years building dashboards in:
Power BI.
Tableau.
Sigma.
ThoughtSpot.
Oracle BI.
Or other BI systems.
It’s unrealistic to discard them just because of AI.
Data agent’s approach is to integrate with them.
OpenAI states it can create or manipulate dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot.
So AI acts as a new analysis entry point, not a replacement for existing BI.
The biggest change is who can start data analysis
Historically, large Data Warehouses were limited to specialized teams.
Sales, marketing, admin, or operations managers seldom accessed data directly—not for lack of questions, but due to tool complexity.
They had to understand tables, joins, metrics, filters, SQL, and dashboards—leading many to just ask data teams for reports.
Data agent replaces that gatekeeper with natural language input
For example, a sales manager could ask:
"Which industries’ pipelines grew fastest this quarter? Compared to last quarter, where is the biggest change coming from?"
A marketer could ask:
"In the past three months, which campaign brought new customers with the best retention?"
A finance worker could ask:
"Which expense categories increased most this quarter? How does that compare to the same period last year?"
An operations manager could ask:
"Is the surge in customer service tickets due to orders, logistics, or product problems?"
This is how Data agent is truly used—not by asking "write me SQL," but by starting with the company’s real questions.
But natural language analysis risks misunderstanding definitions like “revenue”
This is a common pitfall for enterprise data AI.
For instance, “Revenue” might mean:
Order amount?
Completed payments?
Net of refunds?
Including taxes?
Excluding taxes?
Recognized revenue?
If AI doesn’t know the company's official definitions, even 100% accurate SQL can lead to incorrect business conclusions.
Therefore, OpenAI emphasizes Business Context.
Data agent uses existing enterprise Semantic Layers
Simply put, a Semantic Layer clearly documents the company’s true definitions:
What constitutes an active customer?
What defines churn?
How is revenue calculated?
What qualifies as a lead?
How tables join and which calculations follow official company algorithms.
Data agent can leverage definitions maintained via dbt, Databricks Genie Ontology, Snowflake Horizon, GitHub, and existing BI dashboards.
This avoids AI inventing the company’s KPIs on its own.
This is often more important than whether one can write SQL
SQL can be perfectly correct, but if the problem definition is off, answers will still be wrong.
For example, if you ask:
"Which customers are likely to churn?"
Without knowing the company’s exact churn criteria, AI can create beautiful but useless dashboards.
The core of enterprise data AI is not just having access to data, but:
Understanding how the company interprets that data.
Data agent respects existing data access permissions
This point is critical.
If your Snowflake account is limited to viewing Taiwan data and not Japan, Data agent won’t override these restrictions just because it is AI.
OpenAI confirms queries follow the original account’s table, row, and column access controls.
Data agent does not gain a “master key” by connecting to ChatGPT.
Admins control which users can access which connections
For example:
Finance can connect to the Finance Data Warehouse.
Marketing can only access Marketing Data.
Sales reps can view Sales Dashboards but not Payroll, HR, or sensitive customer data.
This aligns with real-world enterprise environments.
It’s not about dumping all data into AI, but about
applying existing data governance to AI.
Insights can lead to actionable next steps
This differentiates Data agent from plain BI chatbots.
For instance, if analysis shows enterprise customer renewals are declining, AI can recommend collaborating with Customer Success, Product, and Sales teams.
Results can be sent via Slack, email, or other connected work apps, or even trigger approved actions.
This creates a complete:
Question → Analysis → Dashboard → Action
workflow.
But AI identified “causes” aren’t always proven causes
For example, AI might find support tickets increased in a month when retention dropped, suggesting “customer service issues might cause churn.”
While plausible, correlation is not causation.
There may be other factors like pricing changes, competitor promos, product bugs, or customer shifts.
The best use of Data agent is not to accept the first “why” answer as fact, but to pursue:
What’s the evidence?
OpenAI emphasizes viewing the evidence behind each finding
Users can follow up, check data sources, compare segments, confirm time frames, and analyze further.
This verification process is more important than AI giving a neat conclusion.
Data agent’s value is in
speeding up validation, not removing it.
A practical example: sales suddenly drop 12%
Assume the company dashboard shows a sudden sales decline this week.
You ask:
Which region?
AI finds it’s mainly the southern region.
You ask:
New or existing customers?
Results show new customers are steady; the drop is from existing customers.
You ask:
Which products?
Concentrated in one product line.
Checking support tickets during the period reveals increased delivery issues for that product.
Only then are you close to a genuine cause worth investigating.
This is better than the AI simply concluding “market demand dropped” upon seeing a revenue fall.
This is why dashboards shouldn't just be pretty
A good dashboard isn’t:
Ten pie charts.
It helps quickly answer:
What changed?
What’s the comparison?
Where are changes concentrated?
What are possible reasons?
What should be investigated next?
If Data agent can preserve this investigative process, it’s much more valuable than just automating chart creation.
OpenAI itself heavily uses this approach
OpenAI shares that almost all product teams and over two-thirds of go-to-market teams use Data agent capabilities to analyze company data themselves.
However, they stress that the prerequisite isn't dumping all data into the model, but rather that data teams have established shared business definitions, access rules, and data safeguards first.
Data agent won’t replace data teams
Instead, it may shift their work focus.
Previously, data teams spent a lot of time pulling numbers, adjusting charts, and adding filters.
In the future, they will likely focus more on defining metrics, governing data quality, building semantic layers, managing access, and verifying sources.
In other words, less report-making, more data governance.
OpenAI has shared real company examples
For example, NTT DATA notes that dashboard licensing costs, technical barriers, and production burdens made it hard for non-engineers to build dashboards.
Using Data agent, sales and corporate functions can now create and update dashboards with natural language.
ServiceTitan used it to identify product adoption differences
ServiceTitan created dashboards comparing usage of AI Sidekick Atlas.
They found customers using Atlas launched campaigns at about three times the rate of non-users.
This insight was used to improve onboarding processes.
AI didn’t just say “Atlas is good” but created verifiable differences from company data.
micro1 case is even more interesting: AI caught errors in existing dashboards
micro1’s operations team rebuilt a performance tracking dashboard via natural language in about 30 minutes.
During this, AI uncovered errors in the original dashboard.
This shows AI doesn’t just automate old reports but prompts reevaluation of previously accepted reporting logic.
But results don’t guarantee all companies will finish in 30 minutes
Different companies have varying:
Data quality.
Metric definitions.
Warehouse structures.
Permissions.
Schemas.
Context.
If data is messy, columns inconsistent, revenue defined differently by departments, or dashboards neglected for years, even the best agent can’t fix governance issues automatically.
Thus, Data agent suits businesses that already have a reliable data foundation but find analysis entry points too slow.
No Snowflake? Small businesses can still benefit
Not necessarily a showstopper.
Data agent can leverage Google Drive, SharePoint, and other connected sources.
If operational data is limited to a few Google Sheets, deploying a full enterprise data stack depends on needs.
Don’t increase complexity just because the tool is powerful.
For very small teams, a clean spreadsheet may suffice.
What kinds of problems suit Data agent?
First: Recurring questions.
Weekly inquiries about sales, retention, costs, or pipeline.
Second: Cross-source data queries.
Combining sales, support, and spending data.
Third: Continuous deep drill-down.
Not looking at just totals but drill down by region, product, customer type.
Fourth: Insights lead to real actions.
Not just yearly reports but timely responses.
For first use, avoid broad requests like "Analyze entire company"
Too big a scope.
Better to pick one question.
For example:
"Compare this month and last month’s renewal rates; identify the customer segment with the biggest change; list data sources, metric definitions, and possible drivers used."
This makes results easier to verify and helps determine AI usability.
Second step: convert analysis into dashboards
If findings prove valuable, then build refreshable dashboards to update weekly.
This approach has a higher success rate than starting with a demand like "Make a full CEO dashboard from the start."
Third step: consider triggering actionable workflows from insights
For example, if a customer group’s renewal risk rises:
First phase: reporting only.
Second phase: notify Customer Success team.
More advanced: create follow-up tasks.
Avoid jumping directly to automatic client emails after analysis, as business actions require judgment beyond data insights.
This is Data agent’s key difference
Past AI data tools often stopped at:
"Help me analyze."
Data agent advances to:
"Help me find data."
"Tell me what changed."
"Let me continue digging evidence."
"Create dashboards."
"Identify stakeholders."
"Send approved results into workflows."
AI starts shifting from just a chart-making assistant to a
Data Investigation Agent.
But the most important factor remains data governance
If the company’s revenue definitions differ across departments, customer IDs are inconsistent between systems, historical data is incomplete, and permissions are chaotic, Data agent can only speed up seeing the mess.
It cannot replace data governance.
Therefore, as AI data analysis matures, the crucial skill will no longer just be SQL ability but
knowing which company metrics are truly authoritative.
So, when first learning about Data agent, remember one thing:
It’s not:
"No need for data teams anymore."
Nor:
"Dashboards can be made by AI without oversight."
The real innovation is enabling business investigations once dependent on data tools, query languages, and BI operations to be triggered directly from a natural language question, leading all the way to evidence, dashboards, and the next steps.
When a company has much data already, what’s often missing is not more numbers but
faster awareness of which numbers need immediate attention.
This is where Data agent truly shines.
Today, advance a little with AI.
Learn an AI skill every day.
Save a bit of time daily.
Improve a little every day.
SasaDaily grows with you.
