In short: according to the maxonline 2026 study, 96% of the CEO names ChatGPT provided for 150 surveyed DACH mid-market companies were completely fabricated — only 3% of the companies had fully correct information. This isn't a bug; it's how generative language models work: they produce plausible text, not verified facts. For business data, that means you need a deterministic layer that validates answers against real data instead of guessing.
96% Wrong Answers — and Companies Don't Notice
Imagine this: you ask ChatGPT for the CEO of a supplier. The answer sounds plausible, is well-phrased, even includes a founding year. Except: it's all fabricated.
That's exactly what the maxonline AI Hallucination Study 2026 systematically tested. Between March and April 2026, it queried 150 DACH mid-market companies (with an emphasis on Austria, supplemented by companies from Germany and Switzerland) across 11 sectors with over 450 standardized prompts — with sobering results:
- 96% incorrect CEO names — of the names ChatGPT actually provided, 96% were entirely fabricated
- 78% wrong founding year — with deviations of up to 160 years
- 68% wrong employee count — sometimes off by a factor of 10
- Only 3% correct company information overall — just 4 of the 150 companies had every queried detail correct
These aren't outliers. This is the norm.
What Exactly Are AI Hallucinations?
The term sounds harmless. It isn't. An AI hallucination occurs when a language model like ChatGPT generates information that is factually incorrect — but sounds convincing. The model doesn't "know" what it doesn't know. It fills gaps with statistically plausible text.
For general knowledge questions, this is often uncritical. For business data, it becomes dangerous:
- A controller checks competitor data and makes decisions based on fabricated numbers
- A sales rep researches a potential customer — and addresses the wrong CEO
- A procurement manager evaluates suppliers based on hallucinated company data
The problem: the answers sound so confident that most people don't question them.
The Scale: Shadow AI in German Mid-Market Companies
According to Bitkom, 41% of German companies use AI tools in 2026. That sounds like controlled adoption — but often it isn't. Microsoft warns: 29% of employees use unauthorized AI agents at work. Without IT department knowledge. Without privacy review. Without quality control of results.
This shadow AI is not just a privacy issue. It's a quality issue. When employees use ChatGPT for business questions and trust the answers, hallucinated "facts" flow into decisions, reports, and customer proposals.
The costs are real: A data breach in Germany costs companies EUR 3.87 million on average, according to IBM — and breaches involving shadow AI cost around USD 670,000 (roughly EUR 600,000) more than incidents without it, according to IBM's global analysis (IBM Cost of a Data Breach Report 2025). And that only accounts for privacy — not the downstream costs of wrong decisions based on false data.
Why ChatGPT Fails at Business Data
This isn't a bug. It's a fundamental design principle. ChatGPT is a generative language model. It was trained to produce convincing-sounding text — not to deliver accurate data. The difference is critical:
Generative AI (ChatGPT, Gemini, Claude):
- Generates text based on probabilities
- Has no access to your current business data
- Cannot distinguish between fact and fiction
- Always answers — even when it doesn't know the answer
Deterministic AI (e.g., oneAgent):
- Calculates results based on real data sources
- Accesses your ERP, CRM, and DWH data directly
- Delivers traceable, reproducible results
- Says "no data available" when no data is available
It's like the difference between someone who googles your question and summarizes the results, and someone who opens your accounting system and runs the numbers.
Direct Comparison: ChatGPT vs. oneAgent
| Criterion | ChatGPT | oneAgent |
|---|---|---|
| Data basis | Training corpus (outdated, incomplete) | Your real business data (live) |
| Answer method | Text generation (probabilistic) | Calculation (deterministic) |
| Hallucination risk | High (96% for company data) | None — calculates or reports missing data |
| Data sources | Manual copy-paste | 550+ connectors (ERP, CRM, DWH, Shopify...) |
| Traceability | No source attribution | Every answer with data source and calculation path |
| Data privacy | Data on US servers | GDPR-compliant, hosted in Frankfurt |
| Data freshness | Training data cutoff | Real-time access to your systems |
| Free trial | No | Yes — with demo data |
What Does "Deterministic" Actually Mean?
Deterministic means: same question + same data = always the same result. No random component. No creative interpretation.
When you ask oneAgent: "What was revenue in Q3 2025?" — oneAgent reads the revenue data from your ERP system, aggregates it according to defined business rules, and delivers the result. Period. No guessing, no hallucinating, no "I think it might be approximately...".
On top of that, an automatic verification layer validates every answer against your actual data and business rules — before you see it. If the data is insufficient or contradictory, the system clearly states that.
The Consequences for Mid-Market Companies
With 41% AI adoption and 29% shadow AI usage in German companies, hallucinated business data is not a fringe problem. It potentially affects every company where employees use ChatGPT for business research.
The question isn't: "Are our employees using ChatGPT?" It's: "Do they trust the answers?"
If the answer is yes, there's a high probability that incorrect information is flowing into your business processes. And unlike a typo in a spreadsheet, AI hallucinations are systematic and hard to detect.
What Should You Do?
Short-term: Build awareness
Inform your teams: ChatGPT is excellent for writing, brainstorming, and summaries. For business data — revenue figures, company data, market data — it's the wrong tool. Not because it's bad, but because it wasn't built for that. Why ChatGPT in the workplace is risky for business data
Mid-term: Break down data silos
The reason employees use ChatGPT for business questions is often: they can't access the data. The ERP is cumbersome, the BI report is outdated, and an SQL query requires the IT department. When you make data accessible, the reason for shadow AI disappears. More on data silos
Long-term: Deploy deterministic AI
Give your teams a tool that answers their real questions — based on real data, in natural language, without SQL skills. That way they get the speed of ChatGPT with the reliability of a BI system. For an overview of available tools, see our comparison 8 AI Analytics Tools Compared 2026.
Frequently Asked Questions
Where does the 96% figure on ChatGPT hallucinations for company data come from?
From the KI-Halluzinations-Studie 2026 (AI Hallucination Study 2026) by agency maxonline Marketing hfw GmbH, published April 7, 2026. Between March and April 2026 it systematically queried 150 DACH mid-market companies across 11 sectors with over 450 standardized prompts. Result: 96% of the CEO names ChatGPT provided were entirely fabricated.
What else did the maxonline study measure?
Besides 96% fabricated CEO names, the study found 78% incorrect founding years (some off by more than 160 years), 68% incorrect employee counts, and fully correct information for only 3% of the 150 companies (4 out of 150).
Is this an isolated case or a fundamental problem with language models?
A fundamental problem. ChatGPT and comparable language models generate text based on probabilities, not verified facts. For well-known large companies with abundant public training material, answers are often correct. For mid-market companies with a small online footprint, the model fills gaps with plausible-sounding but fabricated details.
What's the difference between ChatGPT and deterministic AI like oneAgent?
ChatGPT generates text on a probabilistic basis and has no access to your current business data. oneAgent calculates results directly from your connected data sources (ERP, CRM, DWH) and automatically validates every answer against the actual data and business rules before displaying it. When data is missing, oneAgent says so explicitly instead of guessing.
Conclusion: Don't Trust Any AI That Doesn't Know Your Data
The maxonline study is clear: 96% of the CEO names ChatGPT provided were fabricated — and only 3% of companies were portrayed consistently correctly. That's not a failure — it's the expected outcome when you use a text tool for data analysis.
The alternative is not "no AI." The alternative is the right AI for the right task. Deterministic data analysis instead of generative text production. Real calculation instead of plausible guessing.
oneAgent connects to your 550+ data sources, calculates results deterministically, automatically validates every answer — and your data never leaves your network. GDPR-compliant, hosted in Frankfurt.
Try the difference: free with realistic shop demo data, no credit card required.