Grounded answers: why a knowledge base keeps AI honest

An AI that writes freely almost always sounds convincing. Whether it is right, you cannot tell by reading it. In customer service that is the problem: a wrong answer about a deadline, a tariff or a document costs more than a minute in a queue.

The way out is grounding. The answer comes from your own sources rather than a model's general knowledge, and it shows where it came from. This post explains how that works and why it matters as much for Assist as for bots.

#What "grounded" means

A language model has seen an enormous amount of text in training, but not this month's terms and conditions. Ask it anyway and it fills the gap with whatever is most likely. The result is called a hallucination: fluent, plausible and wrong.

Grounded answers reverse the order. Search comes first, writing second. The model receives the relevant passages from your knowledge base and may answer only from them. The technical name is retrieval-augmented generation, or RAG.

#How an answer is made, in four steps

1. Understand the request. The actual concern is derived from the customer's question or the running conversation, even when it is phrased awkwardly.
2. Find the sources. A semantic search retrieves the passages that match in meaning, not only those that share the same words.
3. Write the answer. The model condenses the passages it found, in the tone of the channel and with nothing added from outside.
4. Show the source. The answer carries a reference to the article it came from. Anyone who wants to check can do so.

The fourth step is the least conspicuous and the most important. An answer with a source can be checked, corrected and improved. An answer without one can only be believed or doubted.

Univerx Knowledge

Question

Where is our data processed?

In EU regions of your choice. Hosting in Germany is planned. The sub-processor list and the processing-agreement documents are available on request.

SourceTrust Center · Data protection

#Why Assist and bots need the same source

Many organisations run on two truths. The chatbot has its own answer texts, the agents have the intranet, and over time the two drift apart. Customers notice at the latest when the bot says something different from the hotline.

A shared knowledge base ends the double upkeep. Univerx Knowledge serves people, bots and self-service alike. Univerx Assist suggests answers to agents during a conversation from the same articles that AI agents and voicebots use.

The difference lies in the human's role. With Assist, an agent reads the suggestion before it leaves the building. With a bot, the machine answers directly. That is precisely why a bot needs stricter grounding, not looser.

#What the knowledge base has to deliver

Grounding is only as good as its sources. A knowledge base that is meant to carry an AI needs a few properties that are easy to neglect in day-to-day work.

The last point closes the loop. When Assist or a bot finds no matching source, that is useful information: it shows which article is missing. Conversation analytics brings such topics to light before they become a pattern.

#Where grounding reaches its limits

Grounding reduces errors; it does not abolish them. There are three limits worth knowing before you switch a bot on.

First, if the source is wrong, the answer will be wrong too. Maintaining the knowledge base remains editorial work with clear ownership.

Second, some questions need data rather than knowledge. "When will my decision arrive?" is not answered by an article but by the case system. That is where AI agents act in connected systems, with permissions and a log.

Third, not every question should be answered automatically. For complaints, hardship cases or unclear legal positions, handing over to a person is the better answer.

#How to recognise good grounding in a test

Before a bot goes live, a simple test with real questions from your daily work pays off. Collect requests from the last few weeks, deliberately including the difficult and badly phrased ones.

• Every answer names a source, and the source actually says what the answer claims.
• For questions without a matching article, the bot does not answer anyway; it asks back or hands over.
• A changed source leads to a changed answer in the next test run.
• Agents adopt Assist suggestions without having to rewrite them first.

If one of these points fails, look at the knowledge base first, not the model. That is good news, because your editors can improve an article the same day.

#Grounding as part of responsible AI

Answers with evidence are more than a quality issue. They belong to the principles by which Univerx designs AI: disclosure as AI, handover at any time, logging of every action and central approval of models and prompts.

The Trust Center explains how these principles fit together under Responsible AI. For the European legal framework, see EU AI Act.

The guide Knowledge management for customer service shows how to build a knowledge base that can carry this.