The same AI model, four data sources, and four completely different answers. Organizations are increasingly using AI agents for credit assessments, supplier screening, and compliance. The key question is not whether AI sounds convincing, but whether the answer is actually correct.
During our recent webinar, Generative AI Expert Lars Wouters asked an AI agent the exact same question four times: provide a report on Dun & Bradstreet B.V., including the number of employees, revenue, and location. With each attempt, the agent was given access to more information: first only the LLM’s training data, then the internet, followed by the D&B Direct+ API, and finally D&B MCP. The question remained unchanged, but the answers differed significantly. Because this involved our own entity, we were able to precisely assess when AI was simply sounding convincing and when the answer was actually accurate.
An AI model without sources hallucinates with complete confidence
The first answer was based solely on training data. It read smoothly and appeared comprehensive, but turned out to be incorrect in several areas. For example, the model provided an outdated office address, incorrectly described Dun & Bradstreet B.V. as a subsidiary of a US publicly listed parent company, used financial figures from four to five years earlier, and estimated the number of employees at 50 to 70.
This is known as hallucination: the model generates a logical and plausible answer that is not based on current and validated facts. Accurate and inaccurate information are presented with the same level of confidence. For a marketing text, this can often still be corrected. For credit decisions, supplier screening, or compliance checks, however, it can lead to incorrect conclusions.
Interesting read: AI sounds convincing. But convincing is not the same as being true
Internet access makes AI responses more complete, but not more reliable
With internet access, the answer became more up to date. The model correctly identified that Dun & Bradstreet in the Benelux is part of Altares. However, the response was still hardly more useful. The number of employees was given as a range of 51 to 200, and the model could not identify an exact registered address.
Business information is spread across websites, registries, and public documents, where reliability and timeliness are not always clear. The model has to determine which source is most authoritative and, when uncertain, tends to choose broad estimates. The issue is therefore not a lack of data, but a lack of reliable business data..
An API provides access to validated data, but only to what has been connected
In the third part of the experiment, the agent was connected to validated D&B data through a traditional API. This resulted in a much more accurate and verifiable business profile. The agent identified, among other things, the correct address, the exact number of employees, accurate registration numbers, and financial information.
At the same time, the API connection highlighted a limitation. Each data block had to be connected separately. When one block was missing, the agent could not use that information. An API therefore gives AI access to reliable data, but only to the information that has been explicitly connected in advance.
D&B MCP combines reliable data with the right context
In the fourth part, the same question was asked through D&B MCP. With a single endpoint, the agent gained access to all available data blocks as well as the context that determines when specific information is relevant.
That is where the value of MCP lies. An AI agent assessing credit risk requires different information from an agent performing a compliance check. By embedding that expertise into the connection, AI can search more effectively, select more relevant data, and work more consistently, without requiring separate integrations or complex software development for every application.
The quality of AI starts with the quality of the data
With direct access to D&B data, a complete and verifiable business profile was created. Every statement could be traced back to a source with a known origin and level of freshness. This is essential for credit assessments, supplier screening, customer due diligence, and compliance, where organizations must be able to explain the basis for their decisions.
AI models are becoming increasingly powerful, but they are also becoming more similar to one another. The outcome of this experiment is therefore especially relevant: the same LLM was used four times, yet produced four different answers. The quality of the result was determined not by the model itself, but by access to the right data and context.
With the D&B.AI Ecosystem, we help organizations connect AI systems with verified business data. During an AI strategy session, we identify together which applications offer the greatest potential and where reliable data can create the most value within your processes.
