Optimisation of generated response content

By using generative AI, responses are created dynamically based on stored information. If there are cases where the AI chatbot does not understand queries correctly, assigns topics incorrectly or provides incomplete responses, various control tools are available. Targeted optimisation ensures that users receive accurate information and that the conversion rate is increased through high-quality responses.

Scenarios at a glance
Enquiries are not understood
Incorrect assignment of intent
Incomplete or poor-quality responses
4.1 Use of instructions
4.2 Adjustment of the persona
4.3 Addition of AI communication rules
How to correct wrong response content

1. Scenarios at a glance

The following table provides an initial guide to the measures that can be taken to address specific problems:

Problem	Possible solution
Enquiry leads to ‘Not understood’	Use AI feedback or add resources
Incorrect topic is recognised	Correct via AI feedback or create a new intent
Answer is incomplete	Provide instructions in the topic or expand resources
Tone is not appropriate	Adjust persona or check AI communication rules
Answer is incorrect in terms of content	Check data source or use specific topic for control

Define use case
The use case of a channel also helps to place enquiries in the correct context by limiting the topics covered. How the area of application is defined is described in this article.

2. Enquiries are not understood

If an enquiry is not understood, optimisation is primarily carried out via the AI playground. Direct feedback on specific enquiries leads to immediate improvements in AI recognition. This article describes how AI feedback is entered.

If the AI feedback already displays the correct topic, even though the widget displayed the ‘not understood’ message, the AI usually lacks the necessary information to generate a response. In this case, additional resources must be added.

3. Incorrect assignment of intent

If the AI chatbot recognises an enquiry but assigns it to the wrong context, a correction is necessary in the AI Playground. The appropriate topic is selected via the AI feedback. The correction takes effect immediately.

If there is no suitable topic for the enquiry yet, there are two options:

Add a new intent: This is recommended if there is a high volume of similar enquiries or if you need to control the response to a topic. Often, AI-generated topic suggestions are already available and can be adopted with just a few clicks. Adding topics is described in this article.
Assign resources to the standard AI agent: If there is a low volume of user enquiries, it is sufficient to add the relevant resource (e.g. a document) to the standard AI agent. This enables responses even without specific topic assignment. How to add resources is described in this article.

When it makes sense to use standard AI agent and what advantages topics and specialised agents have, particularly in the area of analysis and topic management, is described in this article.

4. Incomplete or poor-quality responses

Incomplete answers are often caused by an incomplete data set. Adding additional resources such as PDFs, website URLs or question-answer pairs usually solves the problem. If this is not sufficient, more in-depth control elements are used, as described below.

4.1 Use of instructions

If the resources are complete but the AI weights information incorrectly or omits details, instructions are used. These are specific prompts within a specific intent. How to store instructions is described in this article.

If the resources and instructions are appropriate and the AI still does not generate a correct answer, the answer generation can be optimised in the AI Playground. How AI feedback is applied in the AI Playground via text input is described in this article.

4.2 Adjustment of the persona

The persona defines the functional role and the code of conduct for the AI agent. This includes defining the functional position, for example as a customer service or sales assistant. Stored rules of conduct, such as tone of voice, refraining from speculation, or relying exclusively on verified sources, ensure a consistent persona. This ensures that the AI chatbot communicates in a manner consistent with the brand and remains within its area of competence. How to create a persona is described in this article.

4.3 Addition of AI communication rules

AI communication guidelines are used for global formal requirements. These define, for example:

The use of informal or formal forms of address.
The use of special characters or emojis.
The use of active rather than passive constructions.

This article explains how the communication rules are set up in detail.

5. How to correct wrong response content

In exceptional cases, the AI may provide incorrect statements (hallucinations). This usually occurs due to access to incorrect resources or incorrect intent assignment.

Steps for troubleshooting:

Check intent and agent assignment: Ensure that the enquiry is placed in the correct context (see chapter 3).
Optimise the data basis: Remove inappropriate resources or replace them with precise question-answer pairs (see chapter 4).
Use topic focus: A specialised agent acts as a filter. By assigning exclusive resources to an agent, the AI is forced to generate the answer based solely on this verified data. The use of specialised agents is described in this article.

For particularly critical information (e.g. legal notices or prices), specific topics with editorial response content should always be used to ensure complete control.