Skip to content
Controllingblog

Home❧AI in Controlling❧Analysing large data volumes in the data warehouse with AI

Analysing large data volumes in the data warehouse with AI

What used to be a week of data work, AI now answers in minutes: questions in everyday language, across several systems. What's possible – and where the limits are.

AI in Controlling10 min readUpdated 14 Nov 2026
Illustration: robot assistant with spreadsheet, document and database Match

A data warehouse collects data from accounting, ERP, CRM and payroll in one place. The data is there – but analysing it has so far required SQL skills, a good understanding of the data model and a lot of time. AI is changing exactly that: controllers can ask questions in plain language, and the AI translates them into queries, runs them and explains the result.

What's possible today

  • Questions in everyday language: “Which ten customers had the largest decline in contribution margin last quarter compared with the prior year – and in which product groups?”
  • AI writes SQL or DAX: the query is generated visibly; you can check, save and reuse it.
  • Connecting several sources: via connectors and interfaces the AI accesses the data warehouse, planning data, Excel files or cloud systems and combines them in one analysis.
  • Patterns and anomalies: unusual postings, price outliers, customers with deteriorating payment behaviour.
  • Explaining results: the AI summarises results in text and suggests suitable charts.

How it works technically

An example

Question: “How has the contribution margin ratio per product group developed over the last twelve months, and where has it fallen by more than two percentage points?”

The AI turns this into a query that sums revenue and variable costs per month and product group from the fact table, joins the calendar and product dimensions, calculates the ratio and compares it with the prior-year month:

SELECT p.product_group, k.year_month, SUM(f.revenue) AS revenue, SUM(f.revenue - f.variable_costs) / NULLIF(SUM(f.revenue), 0) AS cm_ratio FROM fact_sales f JOIN dim_product p ON p.product_id = f.product_id JOIN dim_calendar k ON k.date = f.date WHERE k.date >= DATEADD(month, -12, CURRENT_DATE) GROUP BY p.product_group, k.year_month ORDER BY p.product_group, k.year_month;

Example SQL syntax; individual functions differ between databases.

The AI then summarises in which product groups the ratio has fallen and offers to break the cause down further – into price, volume or mix effects, as described in Budget vs. actual analysis.

Highly complex analyses across different connections

AI becomes really powerful when it combines several sources that are otherwise analysed separately:

  • Data warehouse (actual) + planning system (plan) + sales' Excel forecast → one combined variance analysis.
  • Open items + payment history + credit data → early-warning list for payment defaults.
  • Purchase prices + commodity indices + supplier contracts → forecast of material costs.

The prerequisite is a clean foundation: consistent keys (customer, item, account numbers), clear definitions and a data catalogue. Without this groundwork even the best AI compares apples with oranges.

Practical tipStart with a question whose answer you already know – such as last month's revenue according to the monthly report. If the AI result matches, you can trust the connection and definitions. Only then move on to new questions.

Governance: rules that must not be missing

  • Read-only access: the AI may query data but never change it.
  • Respect permissions: the AI only sees what the respective user is allowed to see – for example no salary data.
  • Make queries visible: every answer must be traceable. No number without its query.
  • Approved services: only use AI solutions approved by the company for this data.
  • Humans decide: AI delivers analyses, not decisions.

Limits

  • AI can build queries that are formally correct but wrong in substance – for example forgetting credit notes or cutting off periods incorrectly.
  • With unclear column names the AI guesses. A maintained data catalogue is therefore more important than the AI tool itself.
  • Very large result sets should be condensed by the database, not by the AI.

Conclusion

AI makes the data warehouse more accessible to controllers than ever before: questions in everyday language, queries in seconds, analyses across many sources. Success, however, depends less on the AI than on the foundation – clean data, clear definitions, read-only access and a human who critically checks every result.

Important when using AIBefore uploading, check whether you are allowed to use the data – company policies and data protection law (e.g. the GDPR in the EU) apply here too. Only process personal and confidential data in approved business AI services. Always check AI results against the source data.

Frequently asked questions

Do I need SQL skills to work with AI in the data warehouse?

Not necessarily, but they help a lot with checking. The AI writes the query – you should be able to understand whether it counts and sums the right things.

Is it safe to connect an AI to the data warehouse?

With read-only access, a clean permission concept, approved services and traceable queries the risk can be managed well. IT and data protection should be involved.

What is a semantic layer?

A description of the data in business language: which tables exist, what columns mean and how KPIs are defined. It helps humans and AI use the data correctly.