Analyzing ERP Data with AI: 5 Ways to Query Your ERP in Plain Language

“Which customers ordered significantly less this quarter than last year?” The answer is in your ERP. Getting it today usually takes an export, a pivot table and an hour of work, or a request to the one person in the company who knows where the data lives.

Artificial intelligence promises that you can simply ask such questions, in full sentences, and get an answer in seconds. Interest is high: according to the study “ERP in der Praxis 2026/2027” by Trovarit, which surveys ERP users in Germany, Austria and Switzerland, 57% now consider AI in ERP highly relevant, up from 16% in 2022. Yet only 1.8% use it broadly.

The reason for this gap: there isn’t one solution, but five fundamentally different ways that vary widely in effort, cost, reliability and data protection. This guide compares all five, shows what your ERP system can already do today, and explains how to stop AI from giving you wrong numbers.

Why analyzing ERP data is so tedious

An ERP system is excellent at recording business transactions: orders, invoices, stock movements. It is hardly built for ad-hoc questions. Anyone who wants to know something that isn’t in a standard report almost always ends up in Excel.

Three figures show the scale:

• Only 11% of small businesses in the EU use a BI tool for analysis at all, compared with 69% of large companies (Eurostat, 2025). Among small and mid-sized businesses, exporting to Excel is the norm.
• 52% of ERP users name fast, easy access to information as the main benefit of AI in ERP (Trovarit 2026/2027).
• 25% of German companies that use AI apply it in controlling and accounting, up from 17% the year before (Bitkom, September 2026). At the same time, 26% cite poor result quality and 25% a lack of traceability as obstacles.

That last point is crucial. An AI that answers quickly is worthless if nobody can be sure the number is right. That’s why we compare the five paths not just on convenience but above all on reliability.

The 5 paths at a glance

Not sure which path fits you? The selection guide gives you a reasoned recommendation in two minutes.

Path 1: Your ERP’s built-in AI

Almost all major ERP vendors introduced AI assistants in 2025 and 2026. The advantage is obvious: no additional platform, no interface, and the AI only sees what the user could see without it.

What you should know:

• The features are often still in preview. In Microsoft Dynamics 365 Business Central, for example, both the chat and analysis assist are labeled “production-ready preview”.
• The scope is limited. Business Central’s analysis assist turns questions such as “show vendors by location sorted by number of purchases” into an analysis of the current list. According to Microsoft, however, it cannot join data across tables and explicitly doesn’t provide insights or conclusions of its own.
• Often cloud-only. Business Central’s AI features are only available online, and Joule is designed for the cloud editions of SAP S/4HANA. If you run your ERP on-premise, you’re often left out. How to add AI anyway is covered in our guide to adding AI to your existing ERP with n8n.

For many companies, built-in AI is still the right first step: it often costs little or nothing extra and gets employees used to asking questions of their data.

Path 2: BI tool with AI

BI tools such as Power BI, Qlik and Tableau now have AI assistants that answer plain-language questions and create charts. The key difference from path 1: the AI doesn’t work directly on ERP tables but on a semantic model, which defines exactly what “revenue”, “contribution margin” or “open items” mean in your company.

Power BI Copilot is the obvious choice in Microsoft environments. Its requirements are often underestimated:

• Copilot requires a paid Fabric capacity of F2 or higher or Power BI Premium. According to Microsoft, a Pro license alone isn’t sufficient. The smallest Fabric capacity costs around $260 a month at Microsoft’s list price, more or less depending on region and billing model.
• Microsoft itself stresses that without preparation, Copilot can misinterpret the data and return generic or inaccurate results.

Alternatives: Qlik Answers combines analysis of Qlik apps with documents. Tableau Next with its “Concierge” agent starts at $40 per user per month, according to its pricing page.

Bottom line on path 2: The most reliable path for recurring KPIs and management reporting. But it requires someone to build and maintain the data model. If you don’t use a BI tool yet, you start here with a project, not a switch.

Path 3: AI queries on approved data views

On this path, a language model translates your question into a database query (SQL), runs it and phrases the result as an answer. Specialists call this text-to-SQL. It works with practically any ERP, including older on-premise versions, because it only needs read access to the database.

How good is it today? On the most important benchmark, BIRD, with over 12,000 questions across 95 databases, the best systems answer 83.4% of questions correctly as of October 2026. Human experts reach 93%. On more realistic, multi-step tasks, even top models score much lower: in the interactive version of the benchmark, GPT-5 solves only 8.7 to 17% of tasks.

For an ERP with hundreds of tables and cryptic field names, this means: text-to-SQL without preparation is risky. With preparation, it’s one of the most powerful paths. The three key building blocks:

1. Approved, read-only views instead of raw tables. Create a database view for each area, such as sales, purchasing and inventory, containing only the necessary fields with understandable names. HR data stays out.
2. A dedicated database user with read-only access to these views. That way the AI technically can’t change anything.
3. A KPI glossary that the AI receives with every question. More on that below.

This path suits companies with in-house IT or a service provider and high data protection requirements: if needed, the language model can even run in your own data center, as we show in our local AI vs. cloud AI comparison.

Path 4: AI assistant via interface (MCP)

In 2025 and 2026, the Model Context Protocol (MCP) became the standard for AI assistants to access business systems. Anthropic, OpenAI and Block handed it over to the Linux Foundation in December 2025. An MCP server exposes a system’s functions so that an assistant such as Claude, ChatGPT or Microsoft Copilot can use them.

This is particularly interesting for ERP data because some vendors now offer official MCP servers:

• Microsoft Dynamics 365 Business Central: The official MCP server allows read-only access by default and works with the identity and permissions of the signed-in user, so all access is traceable. Administrators can additionally enable write operations and database queries.
• Xentral: official MCP server in open beta, free during the beta. Read access to customers, items, orders, inventory, KPIs and open items, among others. Permissions are granted per function.
• Odoo 20: According to the September 2026 release notes, the database can be connected via MCP.

For SAP, there is an MCP gateway in SAP Integration Suite. For SAP Business One, according to partner information, only a sample server exists so far. We found no official MCP server for DATEV, abas, weclapp or proALPHA; here, custom development or partner solutions are needed.

For this path, use an AI assistant on a business plan with a data processing agreement and no training on your data. We compare suitable vendors in our article on ChatGPT alternatives for business.

Path 5: Export and AI analysis

The simplest path: export a list from the ERP as an Excel or CSV file and load it into an AI assistant. For one-off analyses this works surprisingly well, if you follow three rules:

1. Business plan only. Private ChatGPT or Claude accounts have no data processing agreement. Customer or HR data doesn’t belong there.
2. Remove personal data first. Replace customer names with customer numbers, remove contacts and employee names.
3. Let it calculate, not estimate. Make sure the assistant analyzes the data with its analysis feature using code, rather than reading numbers from the text. Spot-check totals against the ERP.

The limits: The data is outdated as soon as the export is done, and every new question requires a new export.

Which ERP can do what? An overview

This fits a figure from the 2026 investment report of DSAG, the German-speaking SAP user group: among SAP customers’ AI applications in production, 77% are not from SAP. Many companies don’t wait for their ERP vendor’s features but combine the ERP with vendor-independent solutions.

Why document AI fails with ERP numbers

Many companies already have an AI that answers questions about internal documents. It’s tempting to simply upload ERP exports there too. With numbers, that doesn’t work, and the reason is technical.

Document AI uses retrieval-augmented generation (RAG): it retrieves the text passages most similar to a question and gives them to the language model to formulate the answer.

But the question “What was revenue from customers in Bavaria in Q3?” doesn’t need similar passages; it needs a complete calculation over all matching records. The search only returns some hits, and the language model then “sums up” whatever it sees in the text. The result sounds plausible and is still wrong. A November 2025 study on so-called aggregative questions reaches exactly this conclusion: conventional RAG methods fail at them, while approaches that translate questions into formal queries perform significantly better.

The rule of thumb:

How to prevent wrong numbers

The biggest risk with AI analysis isn’t that the AI fails to answer, but that it answers convincingly wrong. A flawed database query doesn’t throw an error; it simply returns a different number. Five measures make the difference.

1. A KPI glossary for your AI

What is “revenue”? Including or excluding credit notes? By invoice date or delivery date? Net or gross? What’s obvious to your finance team, the AI doesn’t know.

An April 2026 study examined how much a short document with such definitions helps: a glossary of about 4 KB with KPI definitions and terminology rules increased accuracy by 17 to 23 percentage points across three leading language models. The choice of model, by contrast, made hardly any difference.

That’s the most important insight in this article: a good glossary does more than a better model. For each key KPI, record:

• name and synonyms (e.g. “revenue”, “sales”, “turnover”)
• the exact calculation with tables and fields
• what’s excluded (e.g. credit notes, internal transfers, canceled documents)
• which date counts
• an example with the correct result

2. A semantic model instead of raw tables

A semantic model defines KPIs centrally so that every analysis uses the same definition. That’s the principle behind path 2. In its own test, the vendor dbt measured that queries through such a model answered 98 to 100% of test questions correctly, versus 84 to 90% for direct database queries. The test covered only 11 questions and comes from the vendor itself, but the direction matches the glossary study.

3. The AI shows how it calculated

Have the AI show the underlying query or filters with every answer. Experts can see at a glance whether “revenue 2026” really means the fiscal year or accidentally the calendar year.

4. Check plausibility

During rollout, compare every answer with a known figure, such as the total from the monthly report. Keep a list of 20 test questions with correct answers and repeat the test after every update.

5. Take data quality seriously

According to the BARC Data, BI and Analytics Trend Monitor 2026, data quality management is the top priority for data professionals, ahead of AI. Duplicate customer records, inconsistent product groups or missing assignments lead to wrong results even with the best AI.

Security, data protection and works councils

Read-only access: No AI needs write permissions for analysis. Whether built-in feature, MCP server or database view: always start with read-only access.

Inherit permissions: The AI must not see more than the person asking. With built-in AI and Business Central’s official MCP server, that’s ensured because they work with the signed-in user’s permissions. In Power BI, you handle this with row-level security in the data model; with your own database views, via separate views per role.

Exclude HR data: Salaries, sick days and performance data don’t belong in a generally accessible AI analysis. Keep these areas strictly separate.

Data protection: For any external AI service you need a data processing agreement under Art. 28 GDPR. Check in which region data is processed.

Works council (Germany): If the AI analysis can also be used to evaluate employee behavior or performance, such as revenue per sales rep, the works council’s co-determination right under Section 87(1) No. 6 of the German Works Constitution Act applies. It’s enough that the system is objectively capable of it.

30 example questions for your ERP

These questions show what’s possible with AI analysis. Use them to test a vendor or as a starting point for your KPI glossary.

Management

1. How has revenue developed this quarter compared with the same quarter last year, broken down by product group?
2. Which ten customers account for the largest share of revenue, and how has their share changed?
3. What is the order backlog for the next three months?
4. Which product groups have the lowest contribution margin?
5. How many new customers have we won this year, and how much revenue do they bring?

Finance and controlling

6. Which open items are older than 60 days, sorted by amount?
7. What is the average payment time per customer group?
8. Which cost centers are over budget this month?
9. How have material costs per product group developed over the last twelve months?
10. Which credit notes were issued this quarter, and for what reasons?
11. What share of incoming payments used early-payment discounts?

Sales

12. Which customers ordered more than 30% less this quarter than in the same quarter last year?
13. Which quotes from the last 60 days haven’t been answered yet?
14. Which items are frequently ordered together?
15. What is the quote conversion rate per sales region?
16. Which customers haven’t ordered anything in more than six months?
17. How is revenue split between new and existing customers?

Purchasing

18. Which suppliers delivered late most often this year?
19. How have purchase prices for our 20 most important items developed?
20. For which items do we have only one supplier?
21. Which open purchase orders are overdue?
22. What is the purchasing volume per supplier compared with last year?

Warehouse and logistics

23. Which items are below their reorder point?
24. Which items had no movement in the last twelve months?
25. What is the stock coverage of the 50 best-selling items?
26. Which orders couldn’t be fully delivered due to missing stock?
27. How has inventory value developed per warehouse location?

Production and service

28. Which production orders are delayed, and why?
29. Which products had the most complaints?
30. What is the on-time delivery rate per product group this quarter?

Your first AI analysis in four steps

1. Collect questions: Have management, finance, sales and purchasing each write down five questions they regularly deal with. This shows which data is really needed.
2. Create a KPI glossary: Define the ten most important terms from these questions unambiguously, with calculation, exceptions and an example result.
3. Choose a path and start a pilot: Use the selection guide above, start with a small group and one area, such as sales, and test with the collected questions.
4. Verify, then expand: Compare the answers with known figures. Only once the test questions are reliably answered correctly should you open up more areas and users.

Checklist: Is your ERP ready for AI analysis?

Frequently asked questions

Conclusion: define the terms first, then add AI

Querying ERP data in plain language is no longer a vision in 2026. Almost every ERP vendor offers first AI features, official interfaces for AI assistants are becoming standard, and even a simple export can be analyzed in minutes.

The bottleneck isn’t the technology but clarity. An AI can only answer as reliably as your terms are defined. If you create a KPI glossary first, start with read-only access, and verify every answer during rollout, you get a tool that management and finance trust. If you start with the AI directly, you get fast answers, but not reliable ones.