Part 3 of 3 in the series "AI on Business Data — Reliable, Not Just Plausible": Part 1: ChatGPT vs. oneAgent · Part 2: Correct SQL, Wrong Metric · Part 3: Query Your Data Warehouse with AI (this article)
Many companies have invested in their data platform.
They've brought together data from ERP, CRM, shop, marketing, and finance, built a data warehouse, and created dashboards in Power BI or other BI systems.
The data is there.
Yet in customer conversations we keep encountering the same situation:
An employee has a concrete question. The answer is in the data, but the business unit can't retrieve it directly.
So a ticket is created.
An analyst finds the right tables, aligns on the metric, builds the query, and prepares the result.
A few days later, the answer arrives.
Usually the next question follows immediately:
"Why is this different in this region?"
And the process starts over.
The problem is not that the company has too little data.
The problem is access.
The data warehouse knows the answer. But only a few people can ask it the right question.
A data warehouse is not yet self-service
A data warehouse provides:
- a shared data foundation,
- consolidated information,
- historical comparisons,
- controlled access,
- and the basis for reports.
But access often remains technical.
For a new analysis, someone has to know:
- which tables are relevant,
- how they're connected,
- which metric applies,
- which filters must be set,
- and which date column is the right one.
Those skills sit with BI developers, data engineers, and analysts.
For complex analyses, that makes sense.
It becomes a problem when even simple follow-up questions always depend on specialists.
Why dashboards don't answer every question
Dashboards matter.
They're especially good for:
- recurring metrics,
- standardized management reports,
- KPI monitoring,
- and quick overviews.
A dashboard can show, for example:
Revenue dropped twelve percent in May.
That makes the problem visible.
But not yet explained.
Now the questions begin:
- Which products are driving the decline?
- Does it affect specific regions?
- Are new or existing customers affected?
- Have returns increased?
- Have volumes or prices changed?
Building a dedicated dashboard page for every possible follow-up question is neither realistic nor economical.
The dashboard shows the what. oneAgent helps with the why.
Not every question needs a new dashboard
Over time, many companies accumulate a large number of reports:
- similar dashboards for different teams,
- copies with slightly different filters,
- different versions of the same metric,
- and high maintenance effort.
Yet some questions are only relevant once:
"Why did the conversion rate drop last week?"
Others only emerge from a previous answer:
"Which three products explain the difference?"
For such dynamic questions, a dialogue is better suited than a fixed dashboard. What's behind this approach is explained in What is Natural Language BI?
The user phrases their question in natural language and can build directly on the result.
oneAgent does not replace the data warehouse
With new AI solutions, the worry of yet another platform or a new data silo comes up quickly.
oneAgent takes a different approach.
Operational systems
↓
Existing data warehouse or lakehouse
↓
Data model and metric logic
↓
Power BI, reports, and oneAgent
The existing data platform remains the foundation.
oneAgent adds another way in:
- dashboards for standardized reports,
- oneAgent for flexible questions and follow-up analyses.
Investments in data models, views, metrics, and permissions are preserved.
oneAgent doesn't replace the data warehouse. It makes it conversational. How connecting to your data warehouse works in practice — from the first conversation to the pilot — is shown on our Data Warehouse solution page.
How a question gets answered today
A head of sales wants to know:
"Which customers are buying less than last year?"
The typical process:
- She searches the existing dashboards.
- The view she needs is missing.
- She files a request with the BI team.
- The definition is aligned.
- An analyst builds the analysis.
- The result is delivered.
- The next question emerges.
The effort isn't just in the SQL.
It comes from business clarification, data understanding, metric selection, and result validation.
A pure SQL chatbot therefore only solves part of the problem.
How oneAgent changes the process
The user asks directly:
"Which customers are buying at least 20 percent less this year than in the same period last year?"
This becomes a structured analysis:
- Metric: approved revenue
- Dimension: customer
- Period: current year to date
- Comparison: same period last year
- Condition: decline of at least 20 percent
- Permission: own sales territory
If something is missing, oneAgent can ask:
"Should revenue be compared by order date or invoice date?"
Afterwards, the user can immediately ask a follow-up:
"Which product groups explain the decline among the five largest customers?"
A dialogue emerges instead of a series of individual tickets.
Self-service doesn't mean everyone has to learn SQL
Business users know their business questions.
What they often lack is technical access to the data platform.
Self-service should therefore not mean everyone:
- builds data models,
- writes DAX,
- learns SQL,
- or develops their own reports.
Self-service means:
A user can ask a business question without having to master the technical structure behind it.
oneAgent is meant to connect the business unit's question with the existing data logic.
The BI team is not replaced
oneAgent doesn't make data engineers or analysts redundant.
The BI team remains responsible for:
- data integration,
- data quality,
- data models,
- metrics,
- permissions,
- governance,
- and complex analyses.
What changes most is the ticket workload.
Today, BI teams often deal with recurring tasks:
- grouping known metrics differently,
- changing time periods,
- filtering existing data,
- or building similar analyses for more regions.
If business units can ask such questions themselves, the BI team gains time for more valuable work.
oneAgent doesn't abolish the BI team. It reduces the dependency for known data questions.
Existing metrics must continue to apply
If Power BI shows 935,000 euros of net revenue for a period, but oneAgent returns 1,080,000 euros, you don't get self-service.
You get distrust.
That's why an AI solution should not ignore existing definitions. Why this is exactly where the biggest risk lies is shown in Part 2: Correct SQL, Wrong Metric.
The language model may respond flexibly to questions.
But the underlying metric must stay consistent.
The interface may be flexible. The metric logic must not be.
The existing DWH is an advantage
A data warehouse is not an obstacle for oneAgent.
It's often the most important foundation.
Many problems have already been solved there:
- data sources were connected,
- formats were unified,
- history was built up,
- relationships were modeled,
- and business rules were implemented.
oneAgent can build on that groundwork.
The DWH doesn't have to be perfect for this.
For a first pilot, a clearly scoped area that is sufficiently well understood is enough.
How a sensible pilot starts
The biggest mistake is starting too big:
"We want to query all company data via chat right away."
A scoped pilot is better.
1. Pick one business unit
Well suited are areas with recurring questions and high manual effort, such as:
- controlling,
- sales,
- e-commerce,
- marketing,
- or operations.
2. Define a clear data scope
For the start, it's often enough to have:
- selected views,
- an existing data model,
- or a few relevant tables.
3. Define a few central metrics
Five verified metrics are worth more than a hundred theoretically available ones.
Examples:
- net revenue,
- contribution margin,
- order intake,
- return rate,
- plan-actual variance.
4. Use real questions
For example:
- Which customers are buying less than last year?
- Which cost centers are over budget?
- Which products are losing margin?
- Which campaigns generate profitable revenue?
5. Validate together, then expand
Business unit and data owners review the answers. Only then are more metrics, users, and data sources added.
Examples from practice
Controlling
The dashboard shows:
Cost center Marketing: 18 percent over budget.
The controller asks:
"Which cost types explain the variance?"
Then:
"Which suppliers are responsible for it?"
The dashboard provides the starting point. oneAgent supports the follow-up analysis. How AI data analysis speeds up the entire monthly close is shown in AI data analysis in controlling.
E-commerce
Revenue stays stable, margin declines.
oneAgent can help answer questions like:
- Which products are losing contribution margin?
- Where are return costs rising?
- Which campaigns bring revenue but unprofitable customers?
Sales
A head of sales asks:
"Which opportunities stay in one stage unusually long?"
Then:
"Which of them have had no activity for 30 days?"
Follow-up questions like these are usually too individual for a fixed dashboard.
Permissions must apply throughout
Easy access to data must not mean everyone can see everything.
A user must not receive more data via chat than they're entitled to within the company.
This can concern, among other things:
- regions,
- entities,
- customer groups,
- sensitive columns,
- and specific metrics.
The permission must apply at the actual data query — not just in the frontend or in the prompt.
Dashboard and data agent complement each other
Dashboards are good for:
- recurring steering,
- fixed KPI views,
- management reports,
- and quick overviews.
oneAgent is good for:
- spontaneous questions,
- flexible filters,
- root-cause analyses,
- and individual follow-up questions.
A typical flow:
- A dashboard shows an anomaly.
- The user asks oneAgent for the cause.
- They analyze individual dimensions.
- An important result is saved.
- A regularly relevant insight can later flow into a dashboard.
Standardized reporting and flexible analysis complement each other.
Why we build oneAgent
In projects, we see it again and again:
Companies don't necessarily fail because of missing data.
Many already have a data warehouse, dashboards, and specialized teams.
Still, the data only reaches a limited part of the organization.
Access ends where no suitable report exists and a custom query becomes necessary.
That's exactly the gap we want to close.
Not by replacing the existing system.
But by creating a new interface for it:
natural language.
Our goal is for business users to ask questions without learning SQL or DAX — and without opening a ticket for every follow-up question.
Everyone can be a datastar — not because everyone becomes a data engineer, but because everyone gets safe access to the right answers.
Conclusion
A data warehouse can contain all the information you need.
That doesn't yet mean every employee can use it.
Dashboards deliver important, standardized overviews. New and spontaneous questions, however, often lead back into the ticket process.
oneAgent complements the existing BI landscape with flexible access in natural language.
The decisive foundations stay in place:
- data platform,
- metrics,
- relationships,
- permissions,
- and governance.
oneAgent doesn't build a second data world. It makes the existing data world accessible to more people.
Keep reading
That concludes the series. If you started here: Part 1 covers the question "But we already have ChatGPT," and Part 2 shows how technically correct SQL can still return the wrong metric.