AI in customer communication

6 tips on how to perfectly train the AI in ReplyOne

From planning to self-testing: AI learns exactly what you need for your customer service.

17/03/2023

Good customer service is essential for companies today if they want to achieve a high level of customer loyalty and customer satisfaction. To achieve a high first-fix rate or a fast processing time, many companies now rely on smart software with artificial intelligence.

However, just like new employees, an AI cannot deliver the best results from day one, but first needs training to "familiarize" itself with the job.

Training the AI in ReplyOne is particularly easy and can be carried out by the companies themselves. The quality of the learning data is crucial. The AI requires learning data in the form of real customer inquiries. The better these text files are, the more precisely the AI can learn. When training the AI, you can therefore influence how good the answers to your customers will ultimately be and how useful ReplyOne will be for your service team.

With our six tips, you can optimize the AI of ReplyOne, train optimally for your purposes.

Plan the categories of customer inquiries carefully

Categorization in ReplyOne is uniquely precise. The AI assigns incoming messages to the categories you specify.

It all starts with a precise definition of categories: Each category must cover one topic.

If categories are missing, ReplyOne cannot assign the customer messages correctly. The same applies if the categories are not precisely defined or overlap. An example: The topics "Termination" and "Threat of termination" require different responses to your customers. It therefore makes sense to create two different categories for this.

To plan the categories, first take a look at the customer inquiries in your current service activities. Then create a list of the most frequent topics. You can then flexibly change or add to this list at any time. For example, if you have not considered a category. Or if new topics arise.

In ReplyOne, you can set up a new category at any time - and train the AI on the new topic within a very short time.

Once you have defined the categories, you can next select suitable learning data for the AI.

"ReplyOne is definitely a relief; in the past, emails were stored on the server and then assigned manually in a complicated process. The ease of use is great, it saves work afterwards."

Sebastian Groth, Senior Project Manager, Neodigital Versicherungen AG

Select suitable learning data from your customer requests

Good categories are the be-all and end-all. However, the quality of the learning data is also an essential aspect for the success of the project.

The AI should learn from real customer inquiries. Select the learning files very carefully and make sure that the respectivemessage only one topic includes.

Treats a If the customer inquiry covers several topics at once - for example, a complaint and a product inquiry - the training for the AI becomes more challenging. Contradictions and content errors in the texts can also cause the AI to stumble. For example, if a customer writes about product X but means product Y, the AI cannot understand this. This is where people with the relevant product knowledge have an advantage, as they understand it based on the context.

It is precisely this product knowledge that the AI should learn with the help of the learning data and understand the respective context.

Check the learning data

We live and work in an increasingly diverse society. When choosing training texts, therefore, make sure that they reflect all customer groups with their expressions and possible language differences.

This is important because the AI can only learn the language that it finds in the learning files.

In other words, if the training is carried out exclusively with texts written in the language of education, the AI will not be able to decode colloquially formulated queries, for example.

The AI may then assign such requests incorrectly. In the worst case, they end up in the "Other" folder and are only answered after a delay.

How simple can good customer service be?

It's easy!

We'll show you live.
Free. Fast. Customized.

Write training texts for new topics

If you want to create a new category, you will need between 40 and 50 training texts. So many texts, which also have to be of high quality, are sometimes impossible to find in the usual customer inquiries.

Especially for new topics for which there have only been a few genuine customer inquiries to date.

In such cases, you can also write the training texts yourself. There are a few points to bear in mind: Write the texts in natural language and in your customers' choice of words. Also make sure that you do not create any unintentional patterns. This can happen if you write the fictitious training texts all for the same product group. For example, the AI could then link the topic "complaints" with the product group "watch". Consequently, it would assign all emails with the word "watch" to the complaints folder. A human knows that this is nonsense. However, the AI does not understand the texts, but creates rules based on the training data and applies them. It does not question the meaning of the rules.

ReplyOne's AI can do much more than create a list of keywords. It also searches for terms and phrases that appear in similar contexts in the training texts and are used as synonyms, e.g. "necklace" or "jewelry", "jeans" or "trousers", "complaint" or "claim". It recognizes patterns and can take the context into account. It maps these in association rules that describe the relationships between elements. For example, it recognizes that "doesn't work" or "broken" are often used in connection with the category "complaint". Future texts containing these words are then classified as complaints.

Accurately classify the learning data in ReplyOne

Each training text must be assigned to exactly one category. Have the assignment checked by members of your team so that any incorrect assignments can be recognized. If you cannot agree on which category a training text belongs in, then either your categories are incomplete or overlapping. Or the text is not suitable for the training, for example because it deals with several topics or is formulated in a confusing way.

Have you defined categories and collected enough training data? Then you can start a training run.

Use the word cloud and the AI self-test

After a training run, ReplyOne shows you which words have been recognized as relevant for assignment to a category. You can use this word cloud to check whether any incorrect patterns have crept in.

In addition, the software tests itself by treating and classifying the training texts as new emails after training.

Two key figures are important for you: How many training texts are correctly assigned to the categories (Recall) and how many training texts are assigned to an incorrect category (Precisicion)? If these figures are too low, the AI was unable to derive any clear patterns from the training texts. In this case, the categories were not well formed and the training texts could not be clearly assigned. Or the number of training texts was not sufficient.